diff --git a/evaluation_l1/Muennighoff_xstory_cloze/ar/Answer_Given_options/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/ar/Answer_Given_options/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xstory_cloze/ar/Answer_Given_options/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/ar/Answer_Given_options/results.json
diff --git a/evaluation_l1/Muennighoff_xstory_cloze/ar/Choose_Story_Ending/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/ar/Choose_Story_Ending/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xstory_cloze/ar/Choose_Story_Ending/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/ar/Choose_Story_Ending/results.json
diff --git a/evaluation_l1/Muennighoff_xstory_cloze/ar/Generate_Ending/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/ar/Generate_Ending/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xstory_cloze/ar/Generate_Ending/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/ar/Generate_Ending/results.json
diff --git a/evaluation_l1/Muennighoff_xstory_cloze/ar/Novel_Correct_Ending/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/ar/Novel_Correct_Ending/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xstory_cloze/ar/Novel_Correct_Ending/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/ar/Novel_Correct_Ending/results.json
diff --git a/evaluation_l1/Muennighoff_xstory_cloze/ar/Story_Continuation_and_Options/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/ar/Story_Continuation_and_Options/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xstory_cloze/ar/Story_Continuation_and_Options/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/ar/Story_Continuation_and_Options/results.json
diff --git a/evaluation_l1/Muennighoff_xstory_cloze/es/Answer_Given_options/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/es/Answer_Given_options/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xstory_cloze/es/Answer_Given_options/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/es/Answer_Given_options/results.json
diff --git a/evaluation_l1/Muennighoff_xstory_cloze/es/Choose_Story_Ending/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/es/Choose_Story_Ending/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xstory_cloze/es/Choose_Story_Ending/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/es/Choose_Story_Ending/results.json
diff --git a/evaluation_l1/Muennighoff_xstory_cloze/es/Generate_Ending/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/es/Generate_Ending/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xstory_cloze/es/Generate_Ending/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/es/Generate_Ending/results.json
diff --git a/evaluation_l1/Muennighoff_xstory_cloze/es/Novel_Correct_Ending/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/es/Novel_Correct_Ending/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xstory_cloze/es/Novel_Correct_Ending/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/es/Novel_Correct_Ending/results.json
diff --git a/evaluation_l1/Muennighoff_xstory_cloze/es/Story_Continuation_and_Options/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/es/Story_Continuation_and_Options/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xstory_cloze/es/Story_Continuation_and_Options/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/es/Story_Continuation_and_Options/results.json
diff --git a/evaluation_l1/Muennighoff_xstory_cloze/eu/Answer_Given_options/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/eu/Answer_Given_options/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xstory_cloze/eu/Answer_Given_options/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/eu/Answer_Given_options/results.json
diff --git a/evaluation_l1/Muennighoff_xstory_cloze/eu/Choose_Story_Ending/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/eu/Choose_Story_Ending/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xstory_cloze/eu/Choose_Story_Ending/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/eu/Choose_Story_Ending/results.json
diff --git a/evaluation_l1/Muennighoff_xstory_cloze/eu/Generate_Ending/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/eu/Generate_Ending/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xstory_cloze/eu/Generate_Ending/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/eu/Generate_Ending/results.json
diff --git a/evaluation_l1/Muennighoff_xstory_cloze/eu/Novel_Correct_Ending/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/eu/Novel_Correct_Ending/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xstory_cloze/eu/Novel_Correct_Ending/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/eu/Novel_Correct_Ending/results.json
diff --git a/evaluation_l1/Muennighoff_xstory_cloze/eu/Story_Continuation_and_Options/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/eu/Story_Continuation_and_Options/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xstory_cloze/eu/Story_Continuation_and_Options/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/eu/Story_Continuation_and_Options/results.json
diff --git a/evaluation_l1/Muennighoff_xstory_cloze/hi/Answer_Given_options/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/hi/Answer_Given_options/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xstory_cloze/hi/Answer_Given_options/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/hi/Answer_Given_options/results.json
diff --git a/evaluation_l1/Muennighoff_xstory_cloze/hi/Choose_Story_Ending/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/hi/Choose_Story_Ending/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xstory_cloze/hi/Choose_Story_Ending/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/hi/Choose_Story_Ending/results.json
diff --git a/evaluation_l1/Muennighoff_xstory_cloze/hi/Generate_Ending/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/hi/Generate_Ending/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xstory_cloze/hi/Generate_Ending/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/hi/Generate_Ending/results.json
diff --git a/evaluation_l1/Muennighoff_xstory_cloze/hi/Novel_Correct_Ending/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/hi/Novel_Correct_Ending/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xstory_cloze/hi/Novel_Correct_Ending/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/hi/Novel_Correct_Ending/results.json
diff --git a/evaluation_l1/Muennighoff_xstory_cloze/hi/Story_Continuation_and_Options/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/hi/Story_Continuation_and_Options/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xstory_cloze/hi/Story_Continuation_and_Options/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/hi/Story_Continuation_and_Options/results.json
diff --git a/evaluation_l1/Muennighoff_xstory_cloze/id/Answer_Given_options/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/id/Answer_Given_options/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xstory_cloze/id/Answer_Given_options/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/id/Answer_Given_options/results.json
diff --git a/evaluation_l1/Muennighoff_xstory_cloze/id/Choose_Story_Ending/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/id/Choose_Story_Ending/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xstory_cloze/id/Choose_Story_Ending/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/id/Choose_Story_Ending/results.json
diff --git a/evaluation_l1/Muennighoff_xstory_cloze/id/Generate_Ending/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/id/Generate_Ending/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xstory_cloze/id/Generate_Ending/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/id/Generate_Ending/results.json
diff --git a/evaluation_l1/Muennighoff_xstory_cloze/id/Novel_Correct_Ending/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/id/Novel_Correct_Ending/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xstory_cloze/id/Novel_Correct_Ending/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/id/Novel_Correct_Ending/results.json
diff --git a/evaluation_l1/Muennighoff_xstory_cloze/id/Story_Continuation_and_Options/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/id/Story_Continuation_and_Options/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xstory_cloze/id/Story_Continuation_and_Options/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/id/Story_Continuation_and_Options/results.json
diff --git a/evaluation_l1/Muennighoff_xstory_cloze/zh/Answer_Given_options/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/zh/Answer_Given_options/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xstory_cloze/zh/Answer_Given_options/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/zh/Answer_Given_options/results.json
diff --git a/evaluation_l1/Muennighoff_xstory_cloze/zh/Choose_Story_Ending/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/zh/Choose_Story_Ending/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xstory_cloze/zh/Choose_Story_Ending/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/zh/Choose_Story_Ending/results.json
diff --git a/evaluation_l1/Muennighoff_xstory_cloze/zh/Generate_Ending/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/zh/Generate_Ending/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xstory_cloze/zh/Generate_Ending/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/zh/Generate_Ending/results.json
diff --git a/evaluation_l1/Muennighoff_xstory_cloze/zh/Novel_Correct_Ending/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/zh/Novel_Correct_Ending/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xstory_cloze/zh/Novel_Correct_Ending/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/zh/Novel_Correct_Ending/results.json
diff --git a/evaluation_l1/Muennighoff_xstory_cloze/zh/Story_Continuation_and_Options/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/zh/Story_Continuation_and_Options/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xstory_cloze/zh/Story_Continuation_and_Options/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xstory_cloze/zh/Story_Continuation_and_Options/results.json
diff --git a/evaluation_l1/Muennighoff_xwinograd/en/Replace/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/en/Replace/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xwinograd/en/Replace/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/en/Replace/results.json
diff --git a/evaluation_l1/Muennighoff_xwinograd/en/True_or_False/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/en/True_or_False/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xwinograd/en/True_or_False/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/en/True_or_False/results.json
diff --git a/evaluation_l1/Muennighoff_xwinograd/en/does_underscore_refer_to/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/en/does_underscore_refer_to/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xwinograd/en/does_underscore_refer_to/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/en/does_underscore_refer_to/results.json
diff --git a/evaluation_l1/Muennighoff_xwinograd/en/stand_for/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/en/stand_for/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xwinograd/en/stand_for/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/en/stand_for/results.json
diff --git a/evaluation_l1/Muennighoff_xwinograd/en/underscore_refer_to/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/en/underscore_refer_to/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xwinograd/en/underscore_refer_to/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/en/underscore_refer_to/results.json
diff --git a/evaluation_l1/Muennighoff_xwinograd/fr/Replace/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/fr/Replace/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xwinograd/fr/Replace/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/fr/Replace/results.json
diff --git a/evaluation_l1/Muennighoff_xwinograd/fr/True_or_False/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/fr/True_or_False/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xwinograd/fr/True_or_False/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/fr/True_or_False/results.json
diff --git a/evaluation_l1/Muennighoff_xwinograd/fr/does_underscore_refer_to/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/fr/does_underscore_refer_to/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xwinograd/fr/does_underscore_refer_to/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/fr/does_underscore_refer_to/results.json
diff --git a/evaluation_l1/Muennighoff_xwinograd/fr/stand_for/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/fr/stand_for/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xwinograd/fr/stand_for/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/fr/stand_for/results.json
diff --git a/evaluation_l1/Muennighoff_xwinograd/fr/underscore_refer_to/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/fr/underscore_refer_to/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xwinograd/fr/underscore_refer_to/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/fr/underscore_refer_to/results.json
diff --git a/evaluation_l1/Muennighoff_xwinograd/pt/Replace/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/pt/Replace/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xwinograd/pt/Replace/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/pt/Replace/results.json
diff --git a/evaluation_l1/Muennighoff_xwinograd/pt/True_or_False/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/pt/True_or_False/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xwinograd/pt/True_or_False/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/pt/True_or_False/results.json
diff --git a/evaluation_l1/Muennighoff_xwinograd/pt/does_underscore_refer_to/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/pt/does_underscore_refer_to/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xwinograd/pt/does_underscore_refer_to/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/pt/does_underscore_refer_to/results.json
diff --git a/evaluation_l1/Muennighoff_xwinograd/pt/stand_for/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/pt/stand_for/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xwinograd/pt/stand_for/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/pt/stand_for/results.json
diff --git a/evaluation_l1/Muennighoff_xwinograd/pt/underscore_refer_to/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/pt/underscore_refer_to/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xwinograd/pt/underscore_refer_to/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/pt/underscore_refer_to/results.json
diff --git a/evaluation_l1/Muennighoff_xwinograd/zh/Replace/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/zh/Replace/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xwinograd/zh/Replace/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/zh/Replace/results.json
diff --git a/evaluation_l1/Muennighoff_xwinograd/zh/True_or_False/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/zh/True_or_False/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xwinograd/zh/True_or_False/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/zh/True_or_False/results.json
diff --git a/evaluation_l1/Muennighoff_xwinograd/zh/does_underscore_refer_to/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/zh/does_underscore_refer_to/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xwinograd/zh/does_underscore_refer_to/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/zh/does_underscore_refer_to/results.json
diff --git a/evaluation_l1/Muennighoff_xwinograd/zh/stand_for/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/zh/stand_for/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xwinograd/zh/stand_for/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/zh/stand_for/results.json
diff --git a/evaluation_l1/Muennighoff_xwinograd/zh/underscore_refer_to/results.json b/evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/zh/underscore_refer_to/results.json
similarity index 100%
rename from evaluation_l1/Muennighoff_xwinograd/zh/underscore_refer_to/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/Muennighoff_xwinograd/zh/underscore_refer_to/results.json
diff --git a/evaluation_l1/anli/dev_r1/GPT-3_style/results.json b/evaluation_bloomz-1b7/evaluation_l1/anli/dev_r1/GPT-3_style/results.json
similarity index 100%
rename from evaluation_l1/anli/dev_r1/GPT-3_style/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/anli/dev_r1/GPT-3_style/results.json
diff --git a/evaluation_l1/anli/dev_r1/MNLI_crowdsource/results.json b/evaluation_bloomz-1b7/evaluation_l1/anli/dev_r1/MNLI_crowdsource/results.json
similarity index 100%
rename from evaluation_l1/anli/dev_r1/MNLI_crowdsource/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/anli/dev_r1/MNLI_crowdsource/results.json
diff --git a/evaluation_l1/anli/dev_r1/can_we_infer/results.json b/evaluation_bloomz-1b7/evaluation_l1/anli/dev_r1/can_we_infer/results.json
similarity index 100%
rename from evaluation_l1/anli/dev_r1/can_we_infer/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/anli/dev_r1/can_we_infer/results.json
diff --git a/evaluation_l1/anli/dev_r1/guaranteed_possible_impossible/results.json b/evaluation_bloomz-1b7/evaluation_l1/anli/dev_r1/guaranteed_possible_impossible/results.json
similarity index 100%
rename from evaluation_l1/anli/dev_r1/guaranteed_possible_impossible/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/anli/dev_r1/guaranteed_possible_impossible/results.json
diff --git a/evaluation_l1/anli/dev_r1/justified_in_saying/results.json b/evaluation_bloomz-1b7/evaluation_l1/anli/dev_r1/justified_in_saying/results.json
similarity index 100%
rename from evaluation_l1/anli/dev_r1/justified_in_saying/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/anli/dev_r1/justified_in_saying/results.json
diff --git a/evaluation_l1/anli/dev_r2/GPT-3_style/results.json b/evaluation_bloomz-1b7/evaluation_l1/anli/dev_r2/GPT-3_style/results.json
similarity index 100%
rename from evaluation_l1/anli/dev_r2/GPT-3_style/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/anli/dev_r2/GPT-3_style/results.json
diff --git a/evaluation_l1/anli/dev_r2/MNLI_crowdsource/results.json b/evaluation_bloomz-1b7/evaluation_l1/anli/dev_r2/MNLI_crowdsource/results.json
similarity index 100%
rename from evaluation_l1/anli/dev_r2/MNLI_crowdsource/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/anli/dev_r2/MNLI_crowdsource/results.json
diff --git a/evaluation_l1/anli/dev_r2/can_we_infer/results.json b/evaluation_bloomz-1b7/evaluation_l1/anli/dev_r2/can_we_infer/results.json
similarity index 100%
rename from evaluation_l1/anli/dev_r2/can_we_infer/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/anli/dev_r2/can_we_infer/results.json
diff --git a/evaluation_l1/anli/dev_r2/guaranteed_possible_impossible/results.json b/evaluation_bloomz-1b7/evaluation_l1/anli/dev_r2/guaranteed_possible_impossible/results.json
similarity index 100%
rename from evaluation_l1/anli/dev_r2/guaranteed_possible_impossible/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/anli/dev_r2/guaranteed_possible_impossible/results.json
diff --git a/evaluation_l1/anli/dev_r2/justified_in_saying/results.json b/evaluation_bloomz-1b7/evaluation_l1/anli/dev_r2/justified_in_saying/results.json
similarity index 100%
rename from evaluation_l1/anli/dev_r2/justified_in_saying/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/anli/dev_r2/justified_in_saying/results.json
diff --git a/evaluation_l1/anli/dev_r3/GPT-3_style/results.json b/evaluation_bloomz-1b7/evaluation_l1/anli/dev_r3/GPT-3_style/results.json
similarity index 100%
rename from evaluation_l1/anli/dev_r3/GPT-3_style/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/anli/dev_r3/GPT-3_style/results.json
diff --git a/evaluation_l1/anli/dev_r3/MNLI_crowdsource/results.json b/evaluation_bloomz-1b7/evaluation_l1/anli/dev_r3/MNLI_crowdsource/results.json
similarity index 100%
rename from evaluation_l1/anli/dev_r3/MNLI_crowdsource/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/anli/dev_r3/MNLI_crowdsource/results.json
diff --git a/evaluation_l1/anli/dev_r3/can_we_infer/results.json b/evaluation_bloomz-1b7/evaluation_l1/anli/dev_r3/can_we_infer/results.json
similarity index 100%
rename from evaluation_l1/anli/dev_r3/can_we_infer/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/anli/dev_r3/can_we_infer/results.json
diff --git a/evaluation_l1/anli/dev_r3/guaranteed_possible_impossible/results.json b/evaluation_bloomz-1b7/evaluation_l1/anli/dev_r3/guaranteed_possible_impossible/results.json
similarity index 100%
rename from evaluation_l1/anli/dev_r3/guaranteed_possible_impossible/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/anli/dev_r3/guaranteed_possible_impossible/results.json
diff --git a/evaluation_l1/anli/dev_r3/justified_in_saying/results.json b/evaluation_bloomz-1b7/evaluation_l1/anli/dev_r3/justified_in_saying/results.json
similarity index 100%
rename from evaluation_l1/anli/dev_r3/justified_in_saying/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/anli/dev_r3/justified_in_saying/results.json
diff --git a/evaluation_l1/story_cloze/2016/Answer_Given_options/results.json b/evaluation_bloomz-1b7/evaluation_l1/story_cloze/2016/Answer_Given_options/results.json
similarity index 100%
rename from evaluation_l1/story_cloze/2016/Answer_Given_options/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/story_cloze/2016/Answer_Given_options/results.json
diff --git a/evaluation_l1/story_cloze/2016/Choose_Story_Ending/results.json b/evaluation_bloomz-1b7/evaluation_l1/story_cloze/2016/Choose_Story_Ending/results.json
similarity index 100%
rename from evaluation_l1/story_cloze/2016/Choose_Story_Ending/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/story_cloze/2016/Choose_Story_Ending/results.json
diff --git a/evaluation_l1/story_cloze/2016/Generate_Ending/results.json b/evaluation_bloomz-1b7/evaluation_l1/story_cloze/2016/Generate_Ending/results.json
similarity index 100%
rename from evaluation_l1/story_cloze/2016/Generate_Ending/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/story_cloze/2016/Generate_Ending/results.json
diff --git a/evaluation_l1/story_cloze/2016/Novel_Correct_Ending/results.json b/evaluation_bloomz-1b7/evaluation_l1/story_cloze/2016/Novel_Correct_Ending/results.json
similarity index 100%
rename from evaluation_l1/story_cloze/2016/Novel_Correct_Ending/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/story_cloze/2016/Novel_Correct_Ending/results.json
diff --git a/evaluation_l1/story_cloze/2016/Story_Continuation_and_Options/results.json b/evaluation_bloomz-1b7/evaluation_l1/story_cloze/2016/Story_Continuation_and_Options/results.json
similarity index 100%
rename from evaluation_l1/story_cloze/2016/Story_Continuation_and_Options/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/story_cloze/2016/Story_Continuation_and_Options/results.json
diff --git a/evaluation_l1/super_glue/cb/GPT-3_style/results.json b/evaluation_bloomz-1b7/evaluation_l1/super_glue/cb/GPT-3_style/results.json
similarity index 100%
rename from evaluation_l1/super_glue/cb/GPT-3_style/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/super_glue/cb/GPT-3_style/results.json
diff --git a/evaluation_l1/super_glue/cb/MNLI_crowdsource/results.json b/evaluation_bloomz-1b7/evaluation_l1/super_glue/cb/MNLI_crowdsource/results.json
similarity index 100%
rename from evaluation_l1/super_glue/cb/MNLI_crowdsource/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/super_glue/cb/MNLI_crowdsource/results.json
diff --git a/evaluation_l1/super_glue/cb/can_we_infer/results.json b/evaluation_bloomz-1b7/evaluation_l1/super_glue/cb/can_we_infer/results.json
similarity index 100%
rename from evaluation_l1/super_glue/cb/can_we_infer/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/super_glue/cb/can_we_infer/results.json
diff --git a/evaluation_l1/super_glue/cb/guaranteed_possible_impossible/results.json b/evaluation_bloomz-1b7/evaluation_l1/super_glue/cb/guaranteed_possible_impossible/results.json
similarity index 100%
rename from evaluation_l1/super_glue/cb/guaranteed_possible_impossible/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/super_glue/cb/guaranteed_possible_impossible/results.json
diff --git a/evaluation_l1/super_glue/cb/justified_in_saying/results.json b/evaluation_bloomz-1b7/evaluation_l1/super_glue/cb/justified_in_saying/results.json
similarity index 100%
rename from evaluation_l1/super_glue/cb/justified_in_saying/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/super_glue/cb/justified_in_saying/results.json
diff --git a/evaluation_l1/super_glue/copa/C1_or_C2?_premise/results.json b/evaluation_bloomz-1b7/evaluation_l1/super_glue/copa/C1_or_C2?_premise/results.json
similarity index 100%
rename from evaluation_l1/super_glue/copa/C1_or_C2?_premise/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/super_glue/copa/C1_or_C2?_premise/results.json
diff --git a/evaluation_l1/super_glue/copa/best_option/results.json b/evaluation_bloomz-1b7/evaluation_l1/super_glue/copa/best_option/results.json
similarity index 100%
rename from evaluation_l1/super_glue/copa/best_option/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/super_glue/copa/best_option/results.json
diff --git a/evaluation_l1/super_glue/copa/cause_effect/results.json b/evaluation_bloomz-1b7/evaluation_l1/super_glue/copa/cause_effect/results.json
similarity index 100%
rename from evaluation_l1/super_glue/copa/cause_effect/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/super_glue/copa/cause_effect/results.json
diff --git a/evaluation_l1/super_glue/copa/i_am_hesitating/results.json b/evaluation_bloomz-1b7/evaluation_l1/super_glue/copa/i_am_hesitating/results.json
similarity index 100%
rename from evaluation_l1/super_glue/copa/i_am_hesitating/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/super_glue/copa/i_am_hesitating/results.json
diff --git a/evaluation_l1/super_glue/copa/plausible_alternatives/results.json b/evaluation_bloomz-1b7/evaluation_l1/super_glue/copa/plausible_alternatives/results.json
similarity index 100%
rename from evaluation_l1/super_glue/copa/plausible_alternatives/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/super_glue/copa/plausible_alternatives/results.json
diff --git a/evaluation_l1/super_glue/rte/GPT-3_style/results.json b/evaluation_bloomz-1b7/evaluation_l1/super_glue/rte/GPT-3_style/results.json
similarity index 100%
rename from evaluation_l1/super_glue/rte/GPT-3_style/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/super_glue/rte/GPT-3_style/results.json
diff --git a/evaluation_l1/super_glue/rte/MNLI_crowdsource/results.json b/evaluation_bloomz-1b7/evaluation_l1/super_glue/rte/MNLI_crowdsource/results.json
similarity index 100%
rename from evaluation_l1/super_glue/rte/MNLI_crowdsource/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/super_glue/rte/MNLI_crowdsource/results.json
diff --git a/evaluation_l1/super_glue/rte/does_it_follow_that/results.json b/evaluation_bloomz-1b7/evaluation_l1/super_glue/rte/does_it_follow_that/results.json
similarity index 100%
rename from evaluation_l1/super_glue/rte/does_it_follow_that/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/super_glue/rte/does_it_follow_that/results.json
diff --git a/evaluation_l1/super_glue/rte/guaranteed_true/results.json b/evaluation_bloomz-1b7/evaluation_l1/super_glue/rte/guaranteed_true/results.json
similarity index 100%
rename from evaluation_l1/super_glue/rte/guaranteed_true/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/super_glue/rte/guaranteed_true/results.json
diff --git a/evaluation_l1/super_glue/rte/should_assume/results.json b/evaluation_bloomz-1b7/evaluation_l1/super_glue/rte/should_assume/results.json
similarity index 100%
rename from evaluation_l1/super_glue/rte/should_assume/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/super_glue/rte/should_assume/results.json
diff --git a/evaluation_l1/winogrande/winogrande_xl/Replace/results.json b/evaluation_bloomz-1b7/evaluation_l1/winogrande/winogrande_xl/Replace/results.json
similarity index 100%
rename from evaluation_l1/winogrande/winogrande_xl/Replace/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/winogrande/winogrande_xl/Replace/results.json
diff --git a/evaluation_l1/winogrande/winogrande_xl/True_or_False/results.json b/evaluation_bloomz-1b7/evaluation_l1/winogrande/winogrande_xl/True_or_False/results.json
similarity index 100%
rename from evaluation_l1/winogrande/winogrande_xl/True_or_False/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/winogrande/winogrande_xl/True_or_False/results.json
diff --git a/evaluation_l1/winogrande/winogrande_xl/does_underscore_refer_to/results.json b/evaluation_bloomz-1b7/evaluation_l1/winogrande/winogrande_xl/does_underscore_refer_to/results.json
similarity index 100%
rename from evaluation_l1/winogrande/winogrande_xl/does_underscore_refer_to/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/winogrande/winogrande_xl/does_underscore_refer_to/results.json
diff --git a/evaluation_l1/winogrande/winogrande_xl/stand_for/results.json b/evaluation_bloomz-1b7/evaluation_l1/winogrande/winogrande_xl/stand_for/results.json
similarity index 100%
rename from evaluation_l1/winogrande/winogrande_xl/stand_for/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/winogrande/winogrande_xl/stand_for/results.json
diff --git a/evaluation_l1/winogrande/winogrande_xl/underscore_refer_to/results.json b/evaluation_bloomz-1b7/evaluation_l1/winogrande/winogrande_xl/underscore_refer_to/results.json
similarity index 100%
rename from evaluation_l1/winogrande/winogrande_xl/underscore_refer_to/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/winogrande/winogrande_xl/underscore_refer_to/results.json
diff --git a/evaluation_l1/xcopa/id/C1_or_C2?_premise/results.json b/evaluation_bloomz-1b7/evaluation_l1/xcopa/id/C1_or_C2?_premise/results.json
similarity index 100%
rename from evaluation_l1/xcopa/id/C1_or_C2?_premise/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xcopa/id/C1_or_C2?_premise/results.json
diff --git a/evaluation_l1/xcopa/id/best_option/results.json b/evaluation_bloomz-1b7/evaluation_l1/xcopa/id/best_option/results.json
similarity index 100%
rename from evaluation_l1/xcopa/id/best_option/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xcopa/id/best_option/results.json
diff --git a/evaluation_l1/xcopa/id/cause_effect/results.json b/evaluation_bloomz-1b7/evaluation_l1/xcopa/id/cause_effect/results.json
similarity index 100%
rename from evaluation_l1/xcopa/id/cause_effect/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xcopa/id/cause_effect/results.json
diff --git a/evaluation_l1/xcopa/id/i_am_hesitating/results.json b/evaluation_bloomz-1b7/evaluation_l1/xcopa/id/i_am_hesitating/results.json
similarity index 100%
rename from evaluation_l1/xcopa/id/i_am_hesitating/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xcopa/id/i_am_hesitating/results.json
diff --git a/evaluation_l1/xcopa/id/plausible_alternatives/results.json b/evaluation_bloomz-1b7/evaluation_l1/xcopa/id/plausible_alternatives/results.json
similarity index 100%
rename from evaluation_l1/xcopa/id/plausible_alternatives/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xcopa/id/plausible_alternatives/results.json
diff --git a/evaluation_l1/xcopa/sw/C1_or_C2?_premise/results.json b/evaluation_bloomz-1b7/evaluation_l1/xcopa/sw/C1_or_C2?_premise/results.json
similarity index 100%
rename from evaluation_l1/xcopa/sw/C1_or_C2?_premise/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xcopa/sw/C1_or_C2?_premise/results.json
diff --git a/evaluation_l1/xcopa/sw/best_option/results.json b/evaluation_bloomz-1b7/evaluation_l1/xcopa/sw/best_option/results.json
similarity index 100%
rename from evaluation_l1/xcopa/sw/best_option/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xcopa/sw/best_option/results.json
diff --git a/evaluation_l1/xcopa/sw/cause_effect/results.json b/evaluation_bloomz-1b7/evaluation_l1/xcopa/sw/cause_effect/results.json
similarity index 100%
rename from evaluation_l1/xcopa/sw/cause_effect/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xcopa/sw/cause_effect/results.json
diff --git a/evaluation_l1/xcopa/sw/i_am_hesitating/results.json b/evaluation_bloomz-1b7/evaluation_l1/xcopa/sw/i_am_hesitating/results.json
similarity index 100%
rename from evaluation_l1/xcopa/sw/i_am_hesitating/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xcopa/sw/i_am_hesitating/results.json
diff --git a/evaluation_l1/xcopa/sw/plausible_alternatives/results.json b/evaluation_bloomz-1b7/evaluation_l1/xcopa/sw/plausible_alternatives/results.json
similarity index 100%
rename from evaluation_l1/xcopa/sw/plausible_alternatives/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xcopa/sw/plausible_alternatives/results.json
diff --git a/evaluation_l1/xcopa/ta/C1_or_C2?_premise/results.json b/evaluation_bloomz-1b7/evaluation_l1/xcopa/ta/C1_or_C2?_premise/results.json
similarity index 100%
rename from evaluation_l1/xcopa/ta/C1_or_C2?_premise/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xcopa/ta/C1_or_C2?_premise/results.json
diff --git a/evaluation_l1/xcopa/ta/best_option/results.json b/evaluation_bloomz-1b7/evaluation_l1/xcopa/ta/best_option/results.json
similarity index 100%
rename from evaluation_l1/xcopa/ta/best_option/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xcopa/ta/best_option/results.json
diff --git a/evaluation_l1/xcopa/ta/cause_effect/results.json b/evaluation_bloomz-1b7/evaluation_l1/xcopa/ta/cause_effect/results.json
similarity index 100%
rename from evaluation_l1/xcopa/ta/cause_effect/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xcopa/ta/cause_effect/results.json
diff --git a/evaluation_l1/xcopa/ta/i_am_hesitating/results.json b/evaluation_bloomz-1b7/evaluation_l1/xcopa/ta/i_am_hesitating/results.json
similarity index 100%
rename from evaluation_l1/xcopa/ta/i_am_hesitating/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xcopa/ta/i_am_hesitating/results.json
diff --git a/evaluation_l1/xcopa/ta/plausible_alternatives/results.json b/evaluation_bloomz-1b7/evaluation_l1/xcopa/ta/plausible_alternatives/results.json
similarity index 100%
rename from evaluation_l1/xcopa/ta/plausible_alternatives/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xcopa/ta/plausible_alternatives/results.json
diff --git a/evaluation_l1/xcopa/vi/C1_or_C2?_premise/results.json b/evaluation_bloomz-1b7/evaluation_l1/xcopa/vi/C1_or_C2?_premise/results.json
similarity index 100%
rename from evaluation_l1/xcopa/vi/C1_or_C2?_premise/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xcopa/vi/C1_or_C2?_premise/results.json
diff --git a/evaluation_l1/xcopa/vi/best_option/results.json b/evaluation_bloomz-1b7/evaluation_l1/xcopa/vi/best_option/results.json
similarity index 100%
rename from evaluation_l1/xcopa/vi/best_option/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xcopa/vi/best_option/results.json
diff --git a/evaluation_l1/xcopa/vi/cause_effect/results.json b/evaluation_bloomz-1b7/evaluation_l1/xcopa/vi/cause_effect/results.json
similarity index 100%
rename from evaluation_l1/xcopa/vi/cause_effect/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xcopa/vi/cause_effect/results.json
diff --git a/evaluation_l1/xcopa/vi/i_am_hesitating/results.json b/evaluation_bloomz-1b7/evaluation_l1/xcopa/vi/i_am_hesitating/results.json
similarity index 100%
rename from evaluation_l1/xcopa/vi/i_am_hesitating/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xcopa/vi/i_am_hesitating/results.json
diff --git a/evaluation_l1/xcopa/vi/plausible_alternatives/results.json b/evaluation_bloomz-1b7/evaluation_l1/xcopa/vi/plausible_alternatives/results.json
similarity index 100%
rename from evaluation_l1/xcopa/vi/plausible_alternatives/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xcopa/vi/plausible_alternatives/results.json
diff --git a/evaluation_l1/xcopa/zh/C1_or_C2?_premise/results.json b/evaluation_bloomz-1b7/evaluation_l1/xcopa/zh/C1_or_C2?_premise/results.json
similarity index 100%
rename from evaluation_l1/xcopa/zh/C1_or_C2?_premise/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xcopa/zh/C1_or_C2?_premise/results.json
diff --git a/evaluation_l1/xcopa/zh/best_option/results.json b/evaluation_bloomz-1b7/evaluation_l1/xcopa/zh/best_option/results.json
similarity index 100%
rename from evaluation_l1/xcopa/zh/best_option/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xcopa/zh/best_option/results.json
diff --git a/evaluation_l1/xcopa/zh/cause_effect/results.json b/evaluation_bloomz-1b7/evaluation_l1/xcopa/zh/cause_effect/results.json
similarity index 100%
rename from evaluation_l1/xcopa/zh/cause_effect/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xcopa/zh/cause_effect/results.json
diff --git a/evaluation_l1/xcopa/zh/i_am_hesitating/results.json b/evaluation_bloomz-1b7/evaluation_l1/xcopa/zh/i_am_hesitating/results.json
similarity index 100%
rename from evaluation_l1/xcopa/zh/i_am_hesitating/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xcopa/zh/i_am_hesitating/results.json
diff --git a/evaluation_l1/xcopa/zh/plausible_alternatives/results.json b/evaluation_bloomz-1b7/evaluation_l1/xcopa/zh/plausible_alternatives/results.json
similarity index 100%
rename from evaluation_l1/xcopa/zh/plausible_alternatives/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xcopa/zh/plausible_alternatives/results.json
diff --git a/evaluation_l1/xnli/ar/GPT-3_style/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/ar/GPT-3_style/results.json
similarity index 100%
rename from evaluation_l1/xnli/ar/GPT-3_style/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/ar/GPT-3_style/results.json
diff --git a/evaluation_l1/xnli/ar/MNLI_crowdsource/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/ar/MNLI_crowdsource/results.json
similarity index 100%
rename from evaluation_l1/xnli/ar/MNLI_crowdsource/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/ar/MNLI_crowdsource/results.json
diff --git a/evaluation_l1/xnli/ar/can_we_infer/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/ar/can_we_infer/results.json
similarity index 100%
rename from evaluation_l1/xnli/ar/can_we_infer/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/ar/can_we_infer/results.json
diff --git a/evaluation_l1/xnli/ar/guaranteed_possible_impossible/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/ar/guaranteed_possible_impossible/results.json
similarity index 100%
rename from evaluation_l1/xnli/ar/guaranteed_possible_impossible/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/ar/guaranteed_possible_impossible/results.json
diff --git a/evaluation_l1/xnli/ar/justified_in_saying/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/ar/justified_in_saying/results.json
similarity index 100%
rename from evaluation_l1/xnli/ar/justified_in_saying/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/ar/justified_in_saying/results.json
diff --git a/evaluation_l1/xnli/en/GPT-3_style/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/en/GPT-3_style/results.json
similarity index 100%
rename from evaluation_l1/xnli/en/GPT-3_style/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/en/GPT-3_style/results.json
diff --git a/evaluation_l1/xnli/en/MNLI_crowdsource/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/en/MNLI_crowdsource/results.json
similarity index 100%
rename from evaluation_l1/xnli/en/MNLI_crowdsource/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/en/MNLI_crowdsource/results.json
diff --git a/evaluation_l1/xnli/en/can_we_infer/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/en/can_we_infer/results.json
similarity index 100%
rename from evaluation_l1/xnli/en/can_we_infer/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/en/can_we_infer/results.json
diff --git a/evaluation_l1/xnli/en/guaranteed_possible_impossible/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/en/guaranteed_possible_impossible/results.json
similarity index 100%
rename from evaluation_l1/xnli/en/guaranteed_possible_impossible/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/en/guaranteed_possible_impossible/results.json
diff --git a/evaluation_l1/xnli/en/justified_in_saying/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/en/justified_in_saying/results.json
similarity index 100%
rename from evaluation_l1/xnli/en/justified_in_saying/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/en/justified_in_saying/results.json
diff --git a/evaluation_l1/xnli/es/GPT-3_style/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/es/GPT-3_style/results.json
similarity index 100%
rename from evaluation_l1/xnli/es/GPT-3_style/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/es/GPT-3_style/results.json
diff --git a/evaluation_l1/xnli/es/MNLI_crowdsource/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/es/MNLI_crowdsource/results.json
similarity index 100%
rename from evaluation_l1/xnli/es/MNLI_crowdsource/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/es/MNLI_crowdsource/results.json
diff --git a/evaluation_l1/xnli/es/can_we_infer/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/es/can_we_infer/results.json
similarity index 100%
rename from evaluation_l1/xnli/es/can_we_infer/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/es/can_we_infer/results.json
diff --git a/evaluation_l1/xnli/es/guaranteed_possible_impossible/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/es/guaranteed_possible_impossible/results.json
similarity index 100%
rename from evaluation_l1/xnli/es/guaranteed_possible_impossible/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/es/guaranteed_possible_impossible/results.json
diff --git a/evaluation_l1/xnli/es/justified_in_saying/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/es/justified_in_saying/results.json
similarity index 100%
rename from evaluation_l1/xnli/es/justified_in_saying/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/es/justified_in_saying/results.json
diff --git a/evaluation_l1/xnli/fr/GPT-3_style/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/fr/GPT-3_style/results.json
similarity index 100%
rename from evaluation_l1/xnli/fr/GPT-3_style/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/fr/GPT-3_style/results.json
diff --git a/evaluation_l1/xnli/fr/MNLI_crowdsource/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/fr/MNLI_crowdsource/results.json
similarity index 100%
rename from evaluation_l1/xnli/fr/MNLI_crowdsource/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/fr/MNLI_crowdsource/results.json
diff --git a/evaluation_l1/xnli/fr/can_we_infer/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/fr/can_we_infer/results.json
similarity index 100%
rename from evaluation_l1/xnli/fr/can_we_infer/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/fr/can_we_infer/results.json
diff --git a/evaluation_l1/xnli/fr/guaranteed_possible_impossible/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/fr/guaranteed_possible_impossible/results.json
similarity index 100%
rename from evaluation_l1/xnli/fr/guaranteed_possible_impossible/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/fr/guaranteed_possible_impossible/results.json
diff --git a/evaluation_l1/xnli/fr/justified_in_saying/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/fr/justified_in_saying/results.json
similarity index 100%
rename from evaluation_l1/xnli/fr/justified_in_saying/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/fr/justified_in_saying/results.json
diff --git a/evaluation_l1/xnli/hi/GPT-3_style/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/hi/GPT-3_style/results.json
similarity index 100%
rename from evaluation_l1/xnli/hi/GPT-3_style/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/hi/GPT-3_style/results.json
diff --git a/evaluation_l1/xnli/hi/MNLI_crowdsource/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/hi/MNLI_crowdsource/results.json
similarity index 100%
rename from evaluation_l1/xnli/hi/MNLI_crowdsource/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/hi/MNLI_crowdsource/results.json
diff --git a/evaluation_l1/xnli/hi/can_we_infer/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/hi/can_we_infer/results.json
similarity index 100%
rename from evaluation_l1/xnli/hi/can_we_infer/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/hi/can_we_infer/results.json
diff --git a/evaluation_l1/xnli/hi/guaranteed_possible_impossible/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/hi/guaranteed_possible_impossible/results.json
similarity index 100%
rename from evaluation_l1/xnli/hi/guaranteed_possible_impossible/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/hi/guaranteed_possible_impossible/results.json
diff --git a/evaluation_l1/xnli/hi/justified_in_saying/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/hi/justified_in_saying/results.json
similarity index 100%
rename from evaluation_l1/xnli/hi/justified_in_saying/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/hi/justified_in_saying/results.json
diff --git a/evaluation_l1/xnli/sw/GPT-3_style/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/sw/GPT-3_style/results.json
similarity index 100%
rename from evaluation_l1/xnli/sw/GPT-3_style/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/sw/GPT-3_style/results.json
diff --git a/evaluation_l1/xnli/sw/MNLI_crowdsource/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/sw/MNLI_crowdsource/results.json
similarity index 100%
rename from evaluation_l1/xnli/sw/MNLI_crowdsource/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/sw/MNLI_crowdsource/results.json
diff --git a/evaluation_l1/xnli/sw/can_we_infer/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/sw/can_we_infer/results.json
similarity index 100%
rename from evaluation_l1/xnli/sw/can_we_infer/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/sw/can_we_infer/results.json
diff --git a/evaluation_l1/xnli/sw/guaranteed_possible_impossible/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/sw/guaranteed_possible_impossible/results.json
similarity index 100%
rename from evaluation_l1/xnli/sw/guaranteed_possible_impossible/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/sw/guaranteed_possible_impossible/results.json
diff --git a/evaluation_l1/xnli/sw/justified_in_saying/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/sw/justified_in_saying/results.json
similarity index 100%
rename from evaluation_l1/xnli/sw/justified_in_saying/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/sw/justified_in_saying/results.json
diff --git a/evaluation_l1/xnli/ur/GPT-3_style/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/ur/GPT-3_style/results.json
similarity index 100%
rename from evaluation_l1/xnli/ur/GPT-3_style/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/ur/GPT-3_style/results.json
diff --git a/evaluation_l1/xnli/ur/MNLI_crowdsource/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/ur/MNLI_crowdsource/results.json
similarity index 100%
rename from evaluation_l1/xnli/ur/MNLI_crowdsource/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/ur/MNLI_crowdsource/results.json
diff --git a/evaluation_l1/xnli/ur/can_we_infer/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/ur/can_we_infer/results.json
similarity index 100%
rename from evaluation_l1/xnli/ur/can_we_infer/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/ur/can_we_infer/results.json
diff --git a/evaluation_l1/xnli/ur/guaranteed_possible_impossible/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/ur/guaranteed_possible_impossible/results.json
similarity index 100%
rename from evaluation_l1/xnli/ur/guaranteed_possible_impossible/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/ur/guaranteed_possible_impossible/results.json
diff --git a/evaluation_l1/xnli/ur/justified_in_saying/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/ur/justified_in_saying/results.json
similarity index 100%
rename from evaluation_l1/xnli/ur/justified_in_saying/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/ur/justified_in_saying/results.json
diff --git a/evaluation_l1/xnli/vi/GPT-3_style/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/vi/GPT-3_style/results.json
similarity index 100%
rename from evaluation_l1/xnli/vi/GPT-3_style/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/vi/GPT-3_style/results.json
diff --git a/evaluation_l1/xnli/vi/MNLI_crowdsource/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/vi/MNLI_crowdsource/results.json
similarity index 100%
rename from evaluation_l1/xnli/vi/MNLI_crowdsource/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/vi/MNLI_crowdsource/results.json
diff --git a/evaluation_l1/xnli/vi/can_we_infer/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/vi/can_we_infer/results.json
similarity index 100%
rename from evaluation_l1/xnli/vi/can_we_infer/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/vi/can_we_infer/results.json
diff --git a/evaluation_l1/xnli/vi/guaranteed_possible_impossible/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/vi/guaranteed_possible_impossible/results.json
similarity index 100%
rename from evaluation_l1/xnli/vi/guaranteed_possible_impossible/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/vi/guaranteed_possible_impossible/results.json
diff --git a/evaluation_l1/xnli/vi/justified_in_saying/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/vi/justified_in_saying/results.json
similarity index 100%
rename from evaluation_l1/xnli/vi/justified_in_saying/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/vi/justified_in_saying/results.json
diff --git a/evaluation_l1/xnli/zh/GPT-3_style/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/zh/GPT-3_style/results.json
similarity index 100%
rename from evaluation_l1/xnli/zh/GPT-3_style/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/zh/GPT-3_style/results.json
diff --git a/evaluation_l1/xnli/zh/MNLI_crowdsource/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/zh/MNLI_crowdsource/results.json
similarity index 100%
rename from evaluation_l1/xnli/zh/MNLI_crowdsource/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/zh/MNLI_crowdsource/results.json
diff --git a/evaluation_l1/xnli/zh/can_we_infer/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/zh/can_we_infer/results.json
similarity index 100%
rename from evaluation_l1/xnli/zh/can_we_infer/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/zh/can_we_infer/results.json
diff --git a/evaluation_l1/xnli/zh/guaranteed_possible_impossible/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/zh/guaranteed_possible_impossible/results.json
similarity index 100%
rename from evaluation_l1/xnli/zh/guaranteed_possible_impossible/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/zh/guaranteed_possible_impossible/results.json
diff --git a/evaluation_l1/xnli/zh/justified_in_saying/results.json b/evaluation_bloomz-1b7/evaluation_l1/xnli/zh/justified_in_saying/results.json
similarity index 100%
rename from evaluation_l1/xnli/zh/justified_in_saying/results.json
rename to evaluation_bloomz-1b7/evaluation_l1/xnli/zh/justified_in_saying/results.json
diff --git a/evaluation_l2/Muennighoff_xstory_cloze/my/Answer_Given_options/results.json b/evaluation_bloomz-1b7/evaluation_l2/Muennighoff_xstory_cloze/my/Answer_Given_options/results.json
similarity index 100%
rename from evaluation_l2/Muennighoff_xstory_cloze/my/Answer_Given_options/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/Muennighoff_xstory_cloze/my/Answer_Given_options/results.json
diff --git a/evaluation_l2/Muennighoff_xstory_cloze/my/Choose_Story_Ending/results.json b/evaluation_bloomz-1b7/evaluation_l2/Muennighoff_xstory_cloze/my/Choose_Story_Ending/results.json
similarity index 100%
rename from evaluation_l2/Muennighoff_xstory_cloze/my/Choose_Story_Ending/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/Muennighoff_xstory_cloze/my/Choose_Story_Ending/results.json
diff --git a/evaluation_l2/Muennighoff_xstory_cloze/my/Generate_Ending/results.json b/evaluation_bloomz-1b7/evaluation_l2/Muennighoff_xstory_cloze/my/Generate_Ending/results.json
similarity index 100%
rename from evaluation_l2/Muennighoff_xstory_cloze/my/Generate_Ending/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/Muennighoff_xstory_cloze/my/Generate_Ending/results.json
diff --git a/evaluation_l2/Muennighoff_xstory_cloze/my/Novel_Correct_Ending/results.json b/evaluation_bloomz-1b7/evaluation_l2/Muennighoff_xstory_cloze/my/Novel_Correct_Ending/results.json
similarity index 100%
rename from evaluation_l2/Muennighoff_xstory_cloze/my/Novel_Correct_Ending/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/Muennighoff_xstory_cloze/my/Novel_Correct_Ending/results.json
diff --git a/evaluation_l2/Muennighoff_xstory_cloze/my/Story_Continuation_and_Options/results.json b/evaluation_bloomz-1b7/evaluation_l2/Muennighoff_xstory_cloze/my/Story_Continuation_and_Options/results.json
similarity index 100%
rename from evaluation_l2/Muennighoff_xstory_cloze/my/Story_Continuation_and_Options/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/Muennighoff_xstory_cloze/my/Story_Continuation_and_Options/results.json
diff --git a/evaluation_l2/Muennighoff_xstory_cloze/ru/Answer_Given_options/results.json b/evaluation_bloomz-1b7/evaluation_l2/Muennighoff_xstory_cloze/ru/Answer_Given_options/results.json
similarity index 100%
rename from evaluation_l2/Muennighoff_xstory_cloze/ru/Answer_Given_options/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/Muennighoff_xstory_cloze/ru/Answer_Given_options/results.json
diff --git a/evaluation_l2/Muennighoff_xstory_cloze/ru/Choose_Story_Ending/results.json b/evaluation_bloomz-1b7/evaluation_l2/Muennighoff_xstory_cloze/ru/Choose_Story_Ending/results.json
similarity index 100%
rename from evaluation_l2/Muennighoff_xstory_cloze/ru/Choose_Story_Ending/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/Muennighoff_xstory_cloze/ru/Choose_Story_Ending/results.json
diff --git a/evaluation_l2/Muennighoff_xstory_cloze/ru/Generate_Ending/results.json b/evaluation_bloomz-1b7/evaluation_l2/Muennighoff_xstory_cloze/ru/Generate_Ending/results.json
similarity index 100%
rename from evaluation_l2/Muennighoff_xstory_cloze/ru/Generate_Ending/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/Muennighoff_xstory_cloze/ru/Generate_Ending/results.json
diff --git a/evaluation_l2/Muennighoff_xstory_cloze/ru/Novel_Correct_Ending/results.json b/evaluation_bloomz-1b7/evaluation_l2/Muennighoff_xstory_cloze/ru/Novel_Correct_Ending/results.json
similarity index 100%
rename from evaluation_l2/Muennighoff_xstory_cloze/ru/Novel_Correct_Ending/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/Muennighoff_xstory_cloze/ru/Novel_Correct_Ending/results.json
diff --git a/evaluation_l2/Muennighoff_xstory_cloze/ru/Story_Continuation_and_Options/results.json b/evaluation_bloomz-1b7/evaluation_l2/Muennighoff_xstory_cloze/ru/Story_Continuation_and_Options/results.json
similarity index 100%
rename from evaluation_l2/Muennighoff_xstory_cloze/ru/Story_Continuation_and_Options/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/Muennighoff_xstory_cloze/ru/Story_Continuation_and_Options/results.json
diff --git a/evaluation_l2/Muennighoff_xwinograd/jp/Replace/results.json b/evaluation_bloomz-1b7/evaluation_l2/Muennighoff_xwinograd/jp/Replace/results.json
similarity index 100%
rename from evaluation_l2/Muennighoff_xwinograd/jp/Replace/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/Muennighoff_xwinograd/jp/Replace/results.json
diff --git a/evaluation_l2/Muennighoff_xwinograd/jp/True_or_False/results.json b/evaluation_bloomz-1b7/evaluation_l2/Muennighoff_xwinograd/jp/True_or_False/results.json
similarity index 100%
rename from evaluation_l2/Muennighoff_xwinograd/jp/True_or_False/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/Muennighoff_xwinograd/jp/True_or_False/results.json
diff --git a/evaluation_l2/Muennighoff_xwinograd/jp/does_underscore_refer_to/results.json b/evaluation_bloomz-1b7/evaluation_l2/Muennighoff_xwinograd/jp/does_underscore_refer_to/results.json
similarity index 100%
rename from evaluation_l2/Muennighoff_xwinograd/jp/does_underscore_refer_to/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/Muennighoff_xwinograd/jp/does_underscore_refer_to/results.json
diff --git a/evaluation_l2/Muennighoff_xwinograd/jp/stand_for/results.json b/evaluation_bloomz-1b7/evaluation_l2/Muennighoff_xwinograd/jp/stand_for/results.json
similarity index 100%
rename from evaluation_l2/Muennighoff_xwinograd/jp/stand_for/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/Muennighoff_xwinograd/jp/stand_for/results.json
diff --git a/evaluation_l2/Muennighoff_xwinograd/jp/underscore_refer_to/results.json b/evaluation_bloomz-1b7/evaluation_l2/Muennighoff_xwinograd/jp/underscore_refer_to/results.json
similarity index 100%
rename from evaluation_l2/Muennighoff_xwinograd/jp/underscore_refer_to/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/Muennighoff_xwinograd/jp/underscore_refer_to/results.json
diff --git a/evaluation_l2/xcopa/et/C1_or_C2?_premise/results.json b/evaluation_bloomz-1b7/evaluation_l2/xcopa/et/C1_or_C2?_premise/results.json
similarity index 100%
rename from evaluation_l2/xcopa/et/C1_or_C2?_premise/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xcopa/et/C1_or_C2?_premise/results.json
diff --git a/evaluation_l2/xcopa/et/best_option/results.json b/evaluation_bloomz-1b7/evaluation_l2/xcopa/et/best_option/results.json
similarity index 100%
rename from evaluation_l2/xcopa/et/best_option/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xcopa/et/best_option/results.json
diff --git a/evaluation_l2/xcopa/et/cause_effect/results.json b/evaluation_bloomz-1b7/evaluation_l2/xcopa/et/cause_effect/results.json
similarity index 100%
rename from evaluation_l2/xcopa/et/cause_effect/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xcopa/et/cause_effect/results.json
diff --git a/evaluation_l2/xcopa/et/i_am_hesitating/results.json b/evaluation_bloomz-1b7/evaluation_l2/xcopa/et/i_am_hesitating/results.json
similarity index 100%
rename from evaluation_l2/xcopa/et/i_am_hesitating/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xcopa/et/i_am_hesitating/results.json
diff --git a/evaluation_l2/xcopa/et/plausible_alternatives/results.json b/evaluation_bloomz-1b7/evaluation_l2/xcopa/et/plausible_alternatives/results.json
similarity index 100%
rename from evaluation_l2/xcopa/et/plausible_alternatives/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xcopa/et/plausible_alternatives/results.json
diff --git a/evaluation_l2/xcopa/ht/C1_or_C2?_premise/results.json b/evaluation_bloomz-1b7/evaluation_l2/xcopa/ht/C1_or_C2?_premise/results.json
similarity index 100%
rename from evaluation_l2/xcopa/ht/C1_or_C2?_premise/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xcopa/ht/C1_or_C2?_premise/results.json
diff --git a/evaluation_l2/xcopa/ht/best_option/results.json b/evaluation_bloomz-1b7/evaluation_l2/xcopa/ht/best_option/results.json
similarity index 100%
rename from evaluation_l2/xcopa/ht/best_option/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xcopa/ht/best_option/results.json
diff --git a/evaluation_l2/xcopa/ht/cause_effect/results.json b/evaluation_bloomz-1b7/evaluation_l2/xcopa/ht/cause_effect/results.json
similarity index 100%
rename from evaluation_l2/xcopa/ht/cause_effect/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xcopa/ht/cause_effect/results.json
diff --git a/evaluation_l2/xcopa/ht/i_am_hesitating/results.json b/evaluation_bloomz-1b7/evaluation_l2/xcopa/ht/i_am_hesitating/results.json
similarity index 100%
rename from evaluation_l2/xcopa/ht/i_am_hesitating/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xcopa/ht/i_am_hesitating/results.json
diff --git a/evaluation_l2/xcopa/ht/plausible_alternatives/results.json b/evaluation_bloomz-1b7/evaluation_l2/xcopa/ht/plausible_alternatives/results.json
similarity index 100%
rename from evaluation_l2/xcopa/ht/plausible_alternatives/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xcopa/ht/plausible_alternatives/results.json
diff --git a/evaluation_l2/xcopa/it/C1_or_C2?_premise/results.json b/evaluation_bloomz-1b7/evaluation_l2/xcopa/it/C1_or_C2?_premise/results.json
similarity index 100%
rename from evaluation_l2/xcopa/it/C1_or_C2?_premise/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xcopa/it/C1_or_C2?_premise/results.json
diff --git a/evaluation_l2/xcopa/it/best_option/results.json b/evaluation_bloomz-1b7/evaluation_l2/xcopa/it/best_option/results.json
similarity index 100%
rename from evaluation_l2/xcopa/it/best_option/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xcopa/it/best_option/results.json
diff --git a/evaluation_l2/xcopa/it/cause_effect/results.json b/evaluation_bloomz-1b7/evaluation_l2/xcopa/it/cause_effect/results.json
similarity index 100%
rename from evaluation_l2/xcopa/it/cause_effect/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xcopa/it/cause_effect/results.json
diff --git a/evaluation_l2/xcopa/it/i_am_hesitating/results.json b/evaluation_bloomz-1b7/evaluation_l2/xcopa/it/i_am_hesitating/results.json
similarity index 100%
rename from evaluation_l2/xcopa/it/i_am_hesitating/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xcopa/it/i_am_hesitating/results.json
diff --git a/evaluation_l2/xcopa/it/plausible_alternatives/results.json b/evaluation_bloomz-1b7/evaluation_l2/xcopa/it/plausible_alternatives/results.json
similarity index 100%
rename from evaluation_l2/xcopa/it/plausible_alternatives/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xcopa/it/plausible_alternatives/results.json
diff --git a/evaluation_l2/xcopa/qu/C1_or_C2?_premise/results.json b/evaluation_bloomz-1b7/evaluation_l2/xcopa/qu/C1_or_C2?_premise/results.json
similarity index 100%
rename from evaluation_l2/xcopa/qu/C1_or_C2?_premise/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xcopa/qu/C1_or_C2?_premise/results.json
diff --git a/evaluation_l2/xcopa/qu/best_option/results.json b/evaluation_bloomz-1b7/evaluation_l2/xcopa/qu/best_option/results.json
similarity index 100%
rename from evaluation_l2/xcopa/qu/best_option/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xcopa/qu/best_option/results.json
diff --git a/evaluation_l2/xcopa/qu/cause_effect/results.json b/evaluation_bloomz-1b7/evaluation_l2/xcopa/qu/cause_effect/results.json
similarity index 100%
rename from evaluation_l2/xcopa/qu/cause_effect/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xcopa/qu/cause_effect/results.json
diff --git a/evaluation_l2/xcopa/qu/i_am_hesitating/results.json b/evaluation_bloomz-1b7/evaluation_l2/xcopa/qu/i_am_hesitating/results.json
similarity index 100%
rename from evaluation_l2/xcopa/qu/i_am_hesitating/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xcopa/qu/i_am_hesitating/results.json
diff --git a/evaluation_l2/xcopa/qu/plausible_alternatives/results.json b/evaluation_bloomz-1b7/evaluation_l2/xcopa/qu/plausible_alternatives/results.json
similarity index 100%
rename from evaluation_l2/xcopa/qu/plausible_alternatives/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xcopa/qu/plausible_alternatives/results.json
diff --git a/evaluation_l2/xcopa/tr/C1_or_C2?_premise/results.json b/evaluation_bloomz-1b7/evaluation_l2/xcopa/tr/C1_or_C2?_premise/results.json
similarity index 100%
rename from evaluation_l2/xcopa/tr/C1_or_C2?_premise/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xcopa/tr/C1_or_C2?_premise/results.json
diff --git a/evaluation_l2/xcopa/tr/best_option/results.json b/evaluation_bloomz-1b7/evaluation_l2/xcopa/tr/best_option/results.json
similarity index 100%
rename from evaluation_l2/xcopa/tr/best_option/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xcopa/tr/best_option/results.json
diff --git a/evaluation_l2/xcopa/tr/cause_effect/results.json b/evaluation_bloomz-1b7/evaluation_l2/xcopa/tr/cause_effect/results.json
similarity index 100%
rename from evaluation_l2/xcopa/tr/cause_effect/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xcopa/tr/cause_effect/results.json
diff --git a/evaluation_l2/xcopa/tr/i_am_hesitating/results.json b/evaluation_bloomz-1b7/evaluation_l2/xcopa/tr/i_am_hesitating/results.json
similarity index 100%
rename from evaluation_l2/xcopa/tr/i_am_hesitating/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xcopa/tr/i_am_hesitating/results.json
diff --git a/evaluation_l2/xcopa/tr/plausible_alternatives/results.json b/evaluation_bloomz-1b7/evaluation_l2/xcopa/tr/plausible_alternatives/results.json
similarity index 100%
rename from evaluation_l2/xcopa/tr/plausible_alternatives/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xcopa/tr/plausible_alternatives/results.json
diff --git a/evaluation_l2/xnli/bg/GPT-3_style/results.json b/evaluation_bloomz-1b7/evaluation_l2/xnli/bg/GPT-3_style/results.json
similarity index 100%
rename from evaluation_l2/xnli/bg/GPT-3_style/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xnli/bg/GPT-3_style/results.json
diff --git a/evaluation_l2/xnli/bg/MNLI_crowdsource/results.json b/evaluation_bloomz-1b7/evaluation_l2/xnli/bg/MNLI_crowdsource/results.json
similarity index 100%
rename from evaluation_l2/xnli/bg/MNLI_crowdsource/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xnli/bg/MNLI_crowdsource/results.json
diff --git a/evaluation_l2/xnli/bg/can_we_infer/results.json b/evaluation_bloomz-1b7/evaluation_l2/xnli/bg/can_we_infer/results.json
similarity index 100%
rename from evaluation_l2/xnli/bg/can_we_infer/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xnli/bg/can_we_infer/results.json
diff --git a/evaluation_l2/xnli/bg/guaranteed_possible_impossible/results.json b/evaluation_bloomz-1b7/evaluation_l2/xnli/bg/guaranteed_possible_impossible/results.json
similarity index 100%
rename from evaluation_l2/xnli/bg/guaranteed_possible_impossible/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xnli/bg/guaranteed_possible_impossible/results.json
diff --git a/evaluation_l2/xnli/bg/justified_in_saying/results.json b/evaluation_bloomz-1b7/evaluation_l2/xnli/bg/justified_in_saying/results.json
similarity index 100%
rename from evaluation_l2/xnli/bg/justified_in_saying/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xnli/bg/justified_in_saying/results.json
diff --git a/evaluation_l2/xnli/de/GPT-3_style/results.json b/evaluation_bloomz-1b7/evaluation_l2/xnli/de/GPT-3_style/results.json
similarity index 100%
rename from evaluation_l2/xnli/de/GPT-3_style/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xnli/de/GPT-3_style/results.json
diff --git a/evaluation_l2/xnli/de/MNLI_crowdsource/results.json b/evaluation_bloomz-1b7/evaluation_l2/xnli/de/MNLI_crowdsource/results.json
similarity index 100%
rename from evaluation_l2/xnli/de/MNLI_crowdsource/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xnli/de/MNLI_crowdsource/results.json
diff --git a/evaluation_l2/xnli/de/can_we_infer/results.json b/evaluation_bloomz-1b7/evaluation_l2/xnli/de/can_we_infer/results.json
similarity index 100%
rename from evaluation_l2/xnli/de/can_we_infer/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xnli/de/can_we_infer/results.json
diff --git a/evaluation_l2/xnli/de/guaranteed_possible_impossible/results.json b/evaluation_bloomz-1b7/evaluation_l2/xnli/de/guaranteed_possible_impossible/results.json
similarity index 100%
rename from evaluation_l2/xnli/de/guaranteed_possible_impossible/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xnli/de/guaranteed_possible_impossible/results.json
diff --git a/evaluation_l2/xnli/de/justified_in_saying/results.json b/evaluation_bloomz-1b7/evaluation_l2/xnli/de/justified_in_saying/results.json
similarity index 100%
rename from evaluation_l2/xnli/de/justified_in_saying/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xnli/de/justified_in_saying/results.json
diff --git a/evaluation_l2/xnli/el/GPT-3_style/results.json b/evaluation_bloomz-1b7/evaluation_l2/xnli/el/GPT-3_style/results.json
similarity index 100%
rename from evaluation_l2/xnli/el/GPT-3_style/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xnli/el/GPT-3_style/results.json
diff --git a/evaluation_l2/xnli/el/MNLI_crowdsource/results.json b/evaluation_bloomz-1b7/evaluation_l2/xnli/el/MNLI_crowdsource/results.json
similarity index 100%
rename from evaluation_l2/xnli/el/MNLI_crowdsource/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xnli/el/MNLI_crowdsource/results.json
diff --git a/evaluation_l2/xnli/el/can_we_infer/results.json b/evaluation_bloomz-1b7/evaluation_l2/xnli/el/can_we_infer/results.json
similarity index 100%
rename from evaluation_l2/xnli/el/can_we_infer/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xnli/el/can_we_infer/results.json
diff --git a/evaluation_l2/xnli/el/guaranteed_possible_impossible/results.json b/evaluation_bloomz-1b7/evaluation_l2/xnli/el/guaranteed_possible_impossible/results.json
similarity index 100%
rename from evaluation_l2/xnli/el/guaranteed_possible_impossible/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xnli/el/guaranteed_possible_impossible/results.json
diff --git a/evaluation_l2/xnli/el/justified_in_saying/results.json b/evaluation_bloomz-1b7/evaluation_l2/xnli/el/justified_in_saying/results.json
similarity index 100%
rename from evaluation_l2/xnli/el/justified_in_saying/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xnli/el/justified_in_saying/results.json
diff --git a/evaluation_l2/xnli/ru/GPT-3_style/results.json b/evaluation_bloomz-1b7/evaluation_l2/xnli/ru/GPT-3_style/results.json
similarity index 100%
rename from evaluation_l2/xnli/ru/GPT-3_style/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xnli/ru/GPT-3_style/results.json
diff --git a/evaluation_l2/xnli/ru/MNLI_crowdsource/results.json b/evaluation_bloomz-1b7/evaluation_l2/xnli/ru/MNLI_crowdsource/results.json
similarity index 100%
rename from evaluation_l2/xnli/ru/MNLI_crowdsource/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xnli/ru/MNLI_crowdsource/results.json
diff --git a/evaluation_l2/xnli/ru/can_we_infer/results.json b/evaluation_bloomz-1b7/evaluation_l2/xnli/ru/can_we_infer/results.json
similarity index 100%
rename from evaluation_l2/xnli/ru/can_we_infer/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xnli/ru/can_we_infer/results.json
diff --git a/evaluation_l2/xnli/ru/guaranteed_possible_impossible/results.json b/evaluation_bloomz-1b7/evaluation_l2/xnli/ru/guaranteed_possible_impossible/results.json
similarity index 100%
rename from evaluation_l2/xnli/ru/guaranteed_possible_impossible/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xnli/ru/guaranteed_possible_impossible/results.json
diff --git a/evaluation_l2/xnli/ru/justified_in_saying/results.json b/evaluation_bloomz-1b7/evaluation_l2/xnli/ru/justified_in_saying/results.json
similarity index 100%
rename from evaluation_l2/xnli/ru/justified_in_saying/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xnli/ru/justified_in_saying/results.json
diff --git a/evaluation_l2/xnli/th/GPT-3_style/results.json b/evaluation_bloomz-1b7/evaluation_l2/xnli/th/GPT-3_style/results.json
similarity index 100%
rename from evaluation_l2/xnli/th/GPT-3_style/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xnli/th/GPT-3_style/results.json
diff --git a/evaluation_l2/xnli/th/MNLI_crowdsource/results.json b/evaluation_bloomz-1b7/evaluation_l2/xnli/th/MNLI_crowdsource/results.json
similarity index 100%
rename from evaluation_l2/xnli/th/MNLI_crowdsource/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xnli/th/MNLI_crowdsource/results.json
diff --git a/evaluation_l2/xnli/th/can_we_infer/results.json b/evaluation_bloomz-1b7/evaluation_l2/xnli/th/can_we_infer/results.json
similarity index 100%
rename from evaluation_l2/xnli/th/can_we_infer/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xnli/th/can_we_infer/results.json
diff --git a/evaluation_l2/xnli/th/guaranteed_possible_impossible/results.json b/evaluation_bloomz-1b7/evaluation_l2/xnli/th/guaranteed_possible_impossible/results.json
similarity index 100%
rename from evaluation_l2/xnli/th/guaranteed_possible_impossible/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xnli/th/guaranteed_possible_impossible/results.json
diff --git a/evaluation_l2/xnli/th/justified_in_saying/results.json b/evaluation_bloomz-1b7/evaluation_l2/xnli/th/justified_in_saying/results.json
similarity index 100%
rename from evaluation_l2/xnli/th/justified_in_saying/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xnli/th/justified_in_saying/results.json
diff --git a/evaluation_l2/xnli/tr/GPT-3_style/results.json b/evaluation_bloomz-1b7/evaluation_l2/xnli/tr/GPT-3_style/results.json
similarity index 100%
rename from evaluation_l2/xnli/tr/GPT-3_style/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xnli/tr/GPT-3_style/results.json
diff --git a/evaluation_l2/xnli/tr/MNLI_crowdsource/results.json b/evaluation_bloomz-1b7/evaluation_l2/xnli/tr/MNLI_crowdsource/results.json
similarity index 100%
rename from evaluation_l2/xnli/tr/MNLI_crowdsource/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xnli/tr/MNLI_crowdsource/results.json
diff --git a/evaluation_l2/xnli/tr/can_we_infer/results.json b/evaluation_bloomz-1b7/evaluation_l2/xnli/tr/can_we_infer/results.json
similarity index 100%
rename from evaluation_l2/xnli/tr/can_we_infer/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xnli/tr/can_we_infer/results.json
diff --git a/evaluation_l2/xnli/tr/guaranteed_possible_impossible/results.json b/evaluation_bloomz-1b7/evaluation_l2/xnli/tr/guaranteed_possible_impossible/results.json
similarity index 100%
rename from evaluation_l2/xnli/tr/guaranteed_possible_impossible/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xnli/tr/guaranteed_possible_impossible/results.json
diff --git a/evaluation_l2/xnli/tr/justified_in_saying/results.json b/evaluation_bloomz-1b7/evaluation_l2/xnli/tr/justified_in_saying/results.json
similarity index 100%
rename from evaluation_l2/xnli/tr/justified_in_saying/results.json
rename to evaluation_bloomz-1b7/evaluation_l2/xnli/tr/justified_in_saying/results.json
diff --git a/evaluation_val/amazon_reviews_multi/en/prompt_body_title_to_star/results.json b/evaluation_bloomz-1b7/evaluation_val/amazon_reviews_multi/en/prompt_body_title_to_star/results.json
similarity index 100%
rename from evaluation_val/amazon_reviews_multi/en/prompt_body_title_to_star/results.json
rename to evaluation_bloomz-1b7/evaluation_val/amazon_reviews_multi/en/prompt_body_title_to_star/results.json
diff --git a/evaluation_val/amazon_reviews_multi/en/prompt_review_to_star/results.json b/evaluation_bloomz-1b7/evaluation_val/amazon_reviews_multi/en/prompt_review_to_star/results.json
similarity index 100%
rename from evaluation_val/amazon_reviews_multi/en/prompt_review_to_star/results.json
rename to evaluation_bloomz-1b7/evaluation_val/amazon_reviews_multi/en/prompt_review_to_star/results.json
diff --git a/evaluation_val/amazon_reviews_multi/en/prompt_title_to_star/results.json b/evaluation_bloomz-1b7/evaluation_val/amazon_reviews_multi/en/prompt_title_to_star/results.json
similarity index 100%
rename from evaluation_val/amazon_reviews_multi/en/prompt_title_to_star/results.json
rename to evaluation_bloomz-1b7/evaluation_val/amazon_reviews_multi/en/prompt_title_to_star/results.json
diff --git a/evaluation_val/amazon_reviews_multi/es/prompt_body_title_to_star/results.json b/evaluation_bloomz-1b7/evaluation_val/amazon_reviews_multi/es/prompt_body_title_to_star/results.json
similarity index 100%
rename from evaluation_val/amazon_reviews_multi/es/prompt_body_title_to_star/results.json
rename to evaluation_bloomz-1b7/evaluation_val/amazon_reviews_multi/es/prompt_body_title_to_star/results.json
diff --git a/evaluation_val/amazon_reviews_multi/es/prompt_review_to_star/results.json b/evaluation_bloomz-1b7/evaluation_val/amazon_reviews_multi/es/prompt_review_to_star/results.json
similarity index 100%
rename from evaluation_val/amazon_reviews_multi/es/prompt_review_to_star/results.json
rename to evaluation_bloomz-1b7/evaluation_val/amazon_reviews_multi/es/prompt_review_to_star/results.json
diff --git a/evaluation_val/amazon_reviews_multi/es/prompt_title_to_star/results.json b/evaluation_bloomz-1b7/evaluation_val/amazon_reviews_multi/es/prompt_title_to_star/results.json
similarity index 100%
rename from evaluation_val/amazon_reviews_multi/es/prompt_title_to_star/results.json
rename to evaluation_bloomz-1b7/evaluation_val/amazon_reviews_multi/es/prompt_title_to_star/results.json
diff --git a/evaluation_val/amazon_reviews_multi/fr/prompt_body_title_to_star/results.json b/evaluation_bloomz-1b7/evaluation_val/amazon_reviews_multi/fr/prompt_body_title_to_star/results.json
similarity index 100%
rename from evaluation_val/amazon_reviews_multi/fr/prompt_body_title_to_star/results.json
rename to evaluation_bloomz-1b7/evaluation_val/amazon_reviews_multi/fr/prompt_body_title_to_star/results.json
diff --git a/evaluation_val/amazon_reviews_multi/fr/prompt_review_to_star/results.json b/evaluation_bloomz-1b7/evaluation_val/amazon_reviews_multi/fr/prompt_review_to_star/results.json
similarity index 100%
rename from evaluation_val/amazon_reviews_multi/fr/prompt_review_to_star/results.json
rename to evaluation_bloomz-1b7/evaluation_val/amazon_reviews_multi/fr/prompt_review_to_star/results.json
diff --git a/evaluation_val/amazon_reviews_multi/fr/prompt_title_to_star/results.json b/evaluation_bloomz-1b7/evaluation_val/amazon_reviews_multi/fr/prompt_title_to_star/results.json
similarity index 100%
rename from evaluation_val/amazon_reviews_multi/fr/prompt_title_to_star/results.json
rename to evaluation_bloomz-1b7/evaluation_val/amazon_reviews_multi/fr/prompt_title_to_star/results.json
diff --git a/evaluation_val/amazon_reviews_multi/zh/prompt_body_title_to_star/results.json b/evaluation_bloomz-1b7/evaluation_val/amazon_reviews_multi/zh/prompt_body_title_to_star/results.json
similarity index 100%
rename from evaluation_val/amazon_reviews_multi/zh/prompt_body_title_to_star/results.json
rename to evaluation_bloomz-1b7/evaluation_val/amazon_reviews_multi/zh/prompt_body_title_to_star/results.json
diff --git a/evaluation_val/amazon_reviews_multi/zh/prompt_review_to_star/results.json b/evaluation_bloomz-1b7/evaluation_val/amazon_reviews_multi/zh/prompt_review_to_star/results.json
similarity index 100%
rename from evaluation_val/amazon_reviews_multi/zh/prompt_review_to_star/results.json
rename to evaluation_bloomz-1b7/evaluation_val/amazon_reviews_multi/zh/prompt_review_to_star/results.json
diff --git a/evaluation_val/amazon_reviews_multi/zh/prompt_title_to_star/results.json b/evaluation_bloomz-1b7/evaluation_val/amazon_reviews_multi/zh/prompt_title_to_star/results.json
similarity index 100%
rename from evaluation_val/amazon_reviews_multi/zh/prompt_title_to_star/results.json
rename to evaluation_bloomz-1b7/evaluation_val/amazon_reviews_multi/zh/prompt_title_to_star/results.json
diff --git a/evaluation_val/aqua_rat/raw/Answer_questions_from_options/results.json b/evaluation_bloomz-1b7/evaluation_val/aqua_rat/raw/Answer_questions_from_options/results.json
similarity index 100%
rename from evaluation_val/aqua_rat/raw/Answer_questions_from_options/results.json
rename to evaluation_bloomz-1b7/evaluation_val/aqua_rat/raw/Answer_questions_from_options/results.json
diff --git a/evaluation_val/aqua_rat/raw/answer_quiz/results.json b/evaluation_bloomz-1b7/evaluation_val/aqua_rat/raw/answer_quiz/results.json
similarity index 100%
rename from evaluation_val/aqua_rat/raw/answer_quiz/results.json
rename to evaluation_bloomz-1b7/evaluation_val/aqua_rat/raw/answer_quiz/results.json
diff --git a/evaluation_val/aqua_rat/raw/select_the_best_option/results.json b/evaluation_bloomz-1b7/evaluation_val/aqua_rat/raw/select_the_best_option/results.json
similarity index 100%
rename from evaluation_val/aqua_rat/raw/select_the_best_option/results.json
rename to evaluation_bloomz-1b7/evaluation_val/aqua_rat/raw/select_the_best_option/results.json
diff --git a/evaluation_val/art/choose_hypothesis/results.json b/evaluation_bloomz-1b7/evaluation_val/art/choose_hypothesis/results.json
similarity index 100%
rename from evaluation_val/art/choose_hypothesis/results.json
rename to evaluation_bloomz-1b7/evaluation_val/art/choose_hypothesis/results.json
diff --git a/evaluation_val/art/choose_hypothesis_believable/results.json b/evaluation_bloomz-1b7/evaluation_val/art/choose_hypothesis_believable/results.json
similarity index 100%
rename from evaluation_val/art/choose_hypothesis_believable/results.json
rename to evaluation_bloomz-1b7/evaluation_val/art/choose_hypothesis_believable/results.json
diff --git a/evaluation_val/art/choose_hypothesis_desc/results.json b/evaluation_bloomz-1b7/evaluation_val/art/choose_hypothesis_desc/results.json
similarity index 100%
rename from evaluation_val/art/choose_hypothesis_desc/results.json
rename to evaluation_bloomz-1b7/evaluation_val/art/choose_hypothesis_desc/results.json
diff --git a/evaluation_val/art/choose_hypothesis_likely/results.json b/evaluation_bloomz-1b7/evaluation_val/art/choose_hypothesis_likely/results.json
similarity index 100%
rename from evaluation_val/art/choose_hypothesis_likely/results.json
rename to evaluation_bloomz-1b7/evaluation_val/art/choose_hypothesis_likely/results.json
diff --git a/evaluation_val/art/choose_hypothesis_options/results.json b/evaluation_bloomz-1b7/evaluation_val/art/choose_hypothesis_options/results.json
similarity index 100%
rename from evaluation_val/art/choose_hypothesis_options/results.json
rename to evaluation_bloomz-1b7/evaluation_val/art/choose_hypothesis_options/results.json
diff --git a/evaluation_val/banking77/direct_to_which_department/results.json b/evaluation_bloomz-1b7/evaluation_val/banking77/direct_to_which_department/results.json
similarity index 100%
rename from evaluation_val/banking77/direct_to_which_department/results.json
rename to evaluation_bloomz-1b7/evaluation_val/banking77/direct_to_which_department/results.json
diff --git a/evaluation_val/banking77/help_page_topic/results.json b/evaluation_bloomz-1b7/evaluation_val/banking77/help_page_topic/results.json
similarity index 100%
rename from evaluation_val/banking77/help_page_topic/results.json
rename to evaluation_bloomz-1b7/evaluation_val/banking77/help_page_topic/results.json
diff --git a/evaluation_val/banking77/rephrase_as_banking_term/results.json b/evaluation_bloomz-1b7/evaluation_val/banking77/rephrase_as_banking_term/results.json
similarity index 100%
rename from evaluation_val/banking77/rephrase_as_banking_term/results.json
rename to evaluation_bloomz-1b7/evaluation_val/banking77/rephrase_as_banking_term/results.json
diff --git a/evaluation_val/blbooksgenre/title_genre_classifiction/classify/results.json b/evaluation_bloomz-1b7/evaluation_val/blbooksgenre/title_genre_classifiction/classify/results.json
similarity index 100%
rename from evaluation_val/blbooksgenre/title_genre_classifiction/classify/results.json
rename to evaluation_bloomz-1b7/evaluation_val/blbooksgenre/title_genre_classifiction/classify/results.json
diff --git a/evaluation_val/blbooksgenre/title_genre_classifiction/multi-choice/results.json b/evaluation_bloomz-1b7/evaluation_val/blbooksgenre/title_genre_classifiction/multi-choice/results.json
similarity index 100%
rename from evaluation_val/blbooksgenre/title_genre_classifiction/multi-choice/results.json
rename to evaluation_bloomz-1b7/evaluation_val/blbooksgenre/title_genre_classifiction/multi-choice/results.json
diff --git a/evaluation_val/blbooksgenre/title_genre_classifiction/premise_context_first/results.json b/evaluation_bloomz-1b7/evaluation_val/blbooksgenre/title_genre_classifiction/premise_context_first/results.json
similarity index 100%
rename from evaluation_val/blbooksgenre/title_genre_classifiction/premise_context_first/results.json
rename to evaluation_bloomz-1b7/evaluation_val/blbooksgenre/title_genre_classifiction/premise_context_first/results.json
diff --git a/evaluation_val/blimp/adjunct_island/grammatical_between_1_2/results.json b/evaluation_bloomz-1b7/evaluation_val/blimp/adjunct_island/grammatical_between_1_2/results.json
similarity index 100%
rename from evaluation_val/blimp/adjunct_island/grammatical_between_1_2/results.json
rename to evaluation_bloomz-1b7/evaluation_val/blimp/adjunct_island/grammatical_between_1_2/results.json
diff --git a/evaluation_val/blimp/adjunct_island/grammatical_between_A_B/results.json b/evaluation_bloomz-1b7/evaluation_val/blimp/adjunct_island/grammatical_between_A_B/results.json
similarity index 100%
rename from evaluation_val/blimp/adjunct_island/grammatical_between_A_B/results.json
rename to evaluation_bloomz-1b7/evaluation_val/blimp/adjunct_island/grammatical_between_A_B/results.json
diff --git a/evaluation_val/blimp/adjunct_island/grammatical_which_one_1_2/results.json b/evaluation_bloomz-1b7/evaluation_val/blimp/adjunct_island/grammatical_which_one_1_2/results.json
similarity index 100%
rename from evaluation_val/blimp/adjunct_island/grammatical_which_one_1_2/results.json
rename to evaluation_bloomz-1b7/evaluation_val/blimp/adjunct_island/grammatical_which_one_1_2/results.json
diff --git a/evaluation_val/blimp/adjunct_island/single_sentence_bad_yes_no/results.json b/evaluation_bloomz-1b7/evaluation_val/blimp/adjunct_island/single_sentence_bad_yes_no/results.json
similarity index 100%
rename from evaluation_val/blimp/adjunct_island/single_sentence_bad_yes_no/results.json
rename to evaluation_bloomz-1b7/evaluation_val/blimp/adjunct_island/single_sentence_bad_yes_no/results.json
diff --git a/evaluation_val/blimp/adjunct_island/single_sentence_good_yes_no/results.json b/evaluation_bloomz-1b7/evaluation_val/blimp/adjunct_island/single_sentence_good_yes_no/results.json
similarity index 100%
rename from evaluation_val/blimp/adjunct_island/single_sentence_good_yes_no/results.json
rename to evaluation_bloomz-1b7/evaluation_val/blimp/adjunct_island/single_sentence_good_yes_no/results.json
diff --git a/evaluation_val/climate_fever/claim_and_all_supporting_evidences/results.json b/evaluation_bloomz-1b7/evaluation_val/climate_fever/claim_and_all_supporting_evidences/results.json
similarity index 100%
rename from evaluation_val/climate_fever/claim_and_all_supporting_evidences/results.json
rename to evaluation_bloomz-1b7/evaluation_val/climate_fever/claim_and_all_supporting_evidences/results.json
diff --git a/evaluation_val/climate_fever/fifth_evidence_and_claim_itemization/results.json b/evaluation_bloomz-1b7/evaluation_val/climate_fever/fifth_evidence_and_claim_itemization/results.json
similarity index 100%
rename from evaluation_val/climate_fever/fifth_evidence_and_claim_itemization/results.json
rename to evaluation_bloomz-1b7/evaluation_val/climate_fever/fifth_evidence_and_claim_itemization/results.json
diff --git a/evaluation_val/climate_fever/first_evidence_and_claim_itemization/results.json b/evaluation_bloomz-1b7/evaluation_val/climate_fever/first_evidence_and_claim_itemization/results.json
similarity index 100%
rename from evaluation_val/climate_fever/first_evidence_and_claim_itemization/results.json
rename to evaluation_bloomz-1b7/evaluation_val/climate_fever/first_evidence_and_claim_itemization/results.json
diff --git a/evaluation_val/climate_fever/second_evidence_and_claim_itemization/results.json b/evaluation_bloomz-1b7/evaluation_val/climate_fever/second_evidence_and_claim_itemization/results.json
similarity index 100%
rename from evaluation_val/climate_fever/second_evidence_and_claim_itemization/results.json
rename to evaluation_bloomz-1b7/evaluation_val/climate_fever/second_evidence_and_claim_itemization/results.json
diff --git a/evaluation_val/climate_fever/third_evidence_claim_pair/results.json b/evaluation_bloomz-1b7/evaluation_val/climate_fever/third_evidence_claim_pair/results.json
similarity index 100%
rename from evaluation_val/climate_fever/third_evidence_claim_pair/results.json
rename to evaluation_bloomz-1b7/evaluation_val/climate_fever/third_evidence_claim_pair/results.json
diff --git a/evaluation_val/codah/codah/affirmative_instruction_after_sentence_and_choices/results.json b/evaluation_bloomz-1b7/evaluation_val/codah/codah/affirmative_instruction_after_sentence_and_choices/results.json
similarity index 100%
rename from evaluation_val/codah/codah/affirmative_instruction_after_sentence_and_choices/results.json
rename to evaluation_bloomz-1b7/evaluation_val/codah/codah/affirmative_instruction_after_sentence_and_choices/results.json
diff --git a/evaluation_val/codah/codah/affirmative_instruction_before_sentence_and_choices/results.json b/evaluation_bloomz-1b7/evaluation_val/codah/codah/affirmative_instruction_before_sentence_and_choices/results.json
similarity index 100%
rename from evaluation_val/codah/codah/affirmative_instruction_before_sentence_and_choices/results.json
rename to evaluation_bloomz-1b7/evaluation_val/codah/codah/affirmative_instruction_before_sentence_and_choices/results.json
diff --git a/evaluation_val/codah/codah/interrogative_instruction_after_sentence_and_choices/results.json b/evaluation_bloomz-1b7/evaluation_val/codah/codah/interrogative_instruction_after_sentence_and_choices/results.json
similarity index 100%
rename from evaluation_val/codah/codah/interrogative_instruction_after_sentence_and_choices/results.json
rename to evaluation_bloomz-1b7/evaluation_val/codah/codah/interrogative_instruction_after_sentence_and_choices/results.json
diff --git a/evaluation_val/commonsense_qa/answer_given_question_without_options/results.json b/evaluation_bloomz-1b7/evaluation_val/commonsense_qa/answer_given_question_without_options/results.json
similarity index 100%
rename from evaluation_val/commonsense_qa/answer_given_question_without_options/results.json
rename to evaluation_bloomz-1b7/evaluation_val/commonsense_qa/answer_given_question_without_options/results.json
diff --git a/evaluation_val/commonsense_qa/most_suitable_answer/results.json b/evaluation_bloomz-1b7/evaluation_val/commonsense_qa/most_suitable_answer/results.json
similarity index 100%
rename from evaluation_val/commonsense_qa/most_suitable_answer/results.json
rename to evaluation_bloomz-1b7/evaluation_val/commonsense_qa/most_suitable_answer/results.json
diff --git a/evaluation_val/commonsense_qa/question_answering/results.json b/evaluation_bloomz-1b7/evaluation_val/commonsense_qa/question_answering/results.json
similarity index 100%
rename from evaluation_val/commonsense_qa/question_answering/results.json
rename to evaluation_bloomz-1b7/evaluation_val/commonsense_qa/question_answering/results.json
diff --git a/evaluation_val/conv_ai_3/ambiguous/results.json b/evaluation_bloomz-1b7/evaluation_val/conv_ai_3/ambiguous/results.json
similarity index 100%
rename from evaluation_val/conv_ai_3/ambiguous/results.json
rename to evaluation_bloomz-1b7/evaluation_val/conv_ai_3/ambiguous/results.json
diff --git a/evaluation_val/conv_ai_3/clarification_needed/results.json b/evaluation_bloomz-1b7/evaluation_val/conv_ai_3/clarification_needed/results.json
similarity index 100%
rename from evaluation_val/conv_ai_3/clarification_needed/results.json
rename to evaluation_bloomz-1b7/evaluation_val/conv_ai_3/clarification_needed/results.json
diff --git a/evaluation_val/conv_ai_3/directly_answer/results.json b/evaluation_bloomz-1b7/evaluation_val/conv_ai_3/directly_answer/results.json
similarity index 100%
rename from evaluation_val/conv_ai_3/directly_answer/results.json
rename to evaluation_bloomz-1b7/evaluation_val/conv_ai_3/directly_answer/results.json
diff --git a/evaluation_val/conv_ai_3/score_give_number/results.json b/evaluation_bloomz-1b7/evaluation_val/conv_ai_3/score_give_number/results.json
similarity index 100%
rename from evaluation_val/conv_ai_3/score_give_number/results.json
rename to evaluation_bloomz-1b7/evaluation_val/conv_ai_3/score_give_number/results.json
diff --git a/evaluation_val/conv_ai_3/score_how_much/results.json b/evaluation_bloomz-1b7/evaluation_val/conv_ai_3/score_how_much/results.json
similarity index 100%
rename from evaluation_val/conv_ai_3/score_how_much/results.json
rename to evaluation_bloomz-1b7/evaluation_val/conv_ai_3/score_how_much/results.json
diff --git a/evaluation_val/craigslist_bargains/best_deal/results.json b/evaluation_bloomz-1b7/evaluation_val/craigslist_bargains/best_deal/results.json
similarity index 100%
rename from evaluation_val/craigslist_bargains/best_deal/results.json
rename to evaluation_bloomz-1b7/evaluation_val/craigslist_bargains/best_deal/results.json
diff --git a/evaluation_val/craigslist_bargains/good_deal_for_seller/results.json b/evaluation_bloomz-1b7/evaluation_val/craigslist_bargains/good_deal_for_seller/results.json
similarity index 100%
rename from evaluation_val/craigslist_bargains/good_deal_for_seller/results.json
rename to evaluation_bloomz-1b7/evaluation_val/craigslist_bargains/good_deal_for_seller/results.json
diff --git a/evaluation_val/craigslist_bargains/good_deal_for_seller_no_list_price/results.json b/evaluation_bloomz-1b7/evaluation_val/craigslist_bargains/good_deal_for_seller_no_list_price/results.json
similarity index 100%
rename from evaluation_val/craigslist_bargains/good_deal_for_seller_no_list_price/results.json
rename to evaluation_bloomz-1b7/evaluation_val/craigslist_bargains/good_deal_for_seller_no_list_price/results.json
diff --git a/evaluation_val/craigslist_bargains/good_deal_for_seller_no_list_price_implicit/results.json b/evaluation_bloomz-1b7/evaluation_val/craigslist_bargains/good_deal_for_seller_no_list_price_implicit/results.json
similarity index 100%
rename from evaluation_val/craigslist_bargains/good_deal_for_seller_no_list_price_implicit/results.json
rename to evaluation_bloomz-1b7/evaluation_val/craigslist_bargains/good_deal_for_seller_no_list_price_implicit/results.json
diff --git a/evaluation_val/emotion/answer_question_with_emotion_label/results.json b/evaluation_bloomz-1b7/evaluation_val/emotion/answer_question_with_emotion_label/results.json
similarity index 100%
rename from evaluation_val/emotion/answer_question_with_emotion_label/results.json
rename to evaluation_bloomz-1b7/evaluation_val/emotion/answer_question_with_emotion_label/results.json
diff --git a/evaluation_val/emotion/answer_with_class_label/results.json b/evaluation_bloomz-1b7/evaluation_val/emotion/answer_with_class_label/results.json
similarity index 100%
rename from evaluation_val/emotion/answer_with_class_label/results.json
rename to evaluation_bloomz-1b7/evaluation_val/emotion/answer_with_class_label/results.json
diff --git a/evaluation_val/emotion/choose_the_best_emotion_label/results.json b/evaluation_bloomz-1b7/evaluation_val/emotion/choose_the_best_emotion_label/results.json
similarity index 100%
rename from evaluation_val/emotion/choose_the_best_emotion_label/results.json
rename to evaluation_bloomz-1b7/evaluation_val/emotion/choose_the_best_emotion_label/results.json
diff --git a/evaluation_val/emotion/reply_with_emoation_label/results.json b/evaluation_bloomz-1b7/evaluation_val/emotion/reply_with_emoation_label/results.json
similarity index 100%
rename from evaluation_val/emotion/reply_with_emoation_label/results.json
rename to evaluation_bloomz-1b7/evaluation_val/emotion/reply_with_emoation_label/results.json
diff --git a/evaluation_val/financial_phrasebank/sentences_allagree/bullish_neutral_bearish/results.json b/evaluation_bloomz-1b7/evaluation_val/financial_phrasebank/sentences_allagree/bullish_neutral_bearish/results.json
similarity index 100%
rename from evaluation_val/financial_phrasebank/sentences_allagree/bullish_neutral_bearish/results.json
rename to evaluation_bloomz-1b7/evaluation_val/financial_phrasebank/sentences_allagree/bullish_neutral_bearish/results.json
diff --git a/evaluation_val/financial_phrasebank/sentences_allagree/complementary_industries/results.json b/evaluation_bloomz-1b7/evaluation_val/financial_phrasebank/sentences_allagree/complementary_industries/results.json
similarity index 100%
rename from evaluation_val/financial_phrasebank/sentences_allagree/complementary_industries/results.json
rename to evaluation_bloomz-1b7/evaluation_val/financial_phrasebank/sentences_allagree/complementary_industries/results.json
diff --git a/evaluation_val/financial_phrasebank/sentences_allagree/sentiment/results.json b/evaluation_bloomz-1b7/evaluation_val/financial_phrasebank/sentences_allagree/sentiment/results.json
similarity index 100%
rename from evaluation_val/financial_phrasebank/sentences_allagree/sentiment/results.json
rename to evaluation_bloomz-1b7/evaluation_val/financial_phrasebank/sentences_allagree/sentiment/results.json
diff --git a/evaluation_val/financial_phrasebank/sentences_allagree/share_price_option/results.json b/evaluation_bloomz-1b7/evaluation_val/financial_phrasebank/sentences_allagree/share_price_option/results.json
similarity index 100%
rename from evaluation_val/financial_phrasebank/sentences_allagree/share_price_option/results.json
rename to evaluation_bloomz-1b7/evaluation_val/financial_phrasebank/sentences_allagree/share_price_option/results.json
diff --git a/evaluation_val/financial_phrasebank/sentences_allagree/word_comes_to_mind/results.json b/evaluation_bloomz-1b7/evaluation_val/financial_phrasebank/sentences_allagree/word_comes_to_mind/results.json
similarity index 100%
rename from evaluation_val/financial_phrasebank/sentences_allagree/word_comes_to_mind/results.json
rename to evaluation_bloomz-1b7/evaluation_val/financial_phrasebank/sentences_allagree/word_comes_to_mind/results.json
diff --git a/evaluation_val/glue/cola/Following_sentence_acceptable/results.json b/evaluation_bloomz-1b7/evaluation_val/glue/cola/Following_sentence_acceptable/results.json
similarity index 100%
rename from evaluation_val/glue/cola/Following_sentence_acceptable/results.json
rename to evaluation_bloomz-1b7/evaluation_val/glue/cola/Following_sentence_acceptable/results.json
diff --git a/evaluation_val/glue/cola/Make_sense_yes_no/results.json b/evaluation_bloomz-1b7/evaluation_val/glue/cola/Make_sense_yes_no/results.json
similarity index 100%
rename from evaluation_val/glue/cola/Make_sense_yes_no/results.json
rename to evaluation_bloomz-1b7/evaluation_val/glue/cola/Make_sense_yes_no/results.json
diff --git a/evaluation_val/glue/cola/Previous_sentence_acceptable/results.json b/evaluation_bloomz-1b7/evaluation_val/glue/cola/Previous_sentence_acceptable/results.json
similarity index 100%
rename from evaluation_val/glue/cola/Previous_sentence_acceptable/results.json
rename to evaluation_bloomz-1b7/evaluation_val/glue/cola/Previous_sentence_acceptable/results.json
diff --git a/evaluation_val/glue/cola/editing/results.json b/evaluation_bloomz-1b7/evaluation_val/glue/cola/editing/results.json
similarity index 100%
rename from evaluation_val/glue/cola/editing/results.json
rename to evaluation_bloomz-1b7/evaluation_val/glue/cola/editing/results.json
diff --git a/evaluation_val/glue/cola/is_this_correct/results.json b/evaluation_bloomz-1b7/evaluation_val/glue/cola/is_this_correct/results.json
similarity index 100%
rename from evaluation_val/glue/cola/is_this_correct/results.json
rename to evaluation_bloomz-1b7/evaluation_val/glue/cola/is_this_correct/results.json
diff --git a/evaluation_val/glue/sst2/following_positive_negative/results.json b/evaluation_bloomz-1b7/evaluation_val/glue/sst2/following_positive_negative/results.json
similarity index 100%
rename from evaluation_val/glue/sst2/following_positive_negative/results.json
rename to evaluation_bloomz-1b7/evaluation_val/glue/sst2/following_positive_negative/results.json
diff --git a/evaluation_val/glue/sst2/happy_or_mad/results.json b/evaluation_bloomz-1b7/evaluation_val/glue/sst2/happy_or_mad/results.json
similarity index 100%
rename from evaluation_val/glue/sst2/happy_or_mad/results.json
rename to evaluation_bloomz-1b7/evaluation_val/glue/sst2/happy_or_mad/results.json
diff --git a/evaluation_val/glue/sst2/positive_negative_after/results.json b/evaluation_bloomz-1b7/evaluation_val/glue/sst2/positive_negative_after/results.json
similarity index 100%
rename from evaluation_val/glue/sst2/positive_negative_after/results.json
rename to evaluation_bloomz-1b7/evaluation_val/glue/sst2/positive_negative_after/results.json
diff --git a/evaluation_val/glue/sst2/review/results.json b/evaluation_bloomz-1b7/evaluation_val/glue/sst2/review/results.json
similarity index 100%
rename from evaluation_val/glue/sst2/review/results.json
rename to evaluation_bloomz-1b7/evaluation_val/glue/sst2/review/results.json
diff --git a/evaluation_val/glue/sst2/said/results.json b/evaluation_bloomz-1b7/evaluation_val/glue/sst2/said/results.json
similarity index 100%
rename from evaluation_val/glue/sst2/said/results.json
rename to evaluation_bloomz-1b7/evaluation_val/glue/sst2/said/results.json
diff --git a/evaluation_val/head_qa/en/multiple_choice_a_and_q_en/results.json b/evaluation_bloomz-1b7/evaluation_val/head_qa/en/multiple_choice_a_and_q_en/results.json
similarity index 100%
rename from evaluation_val/head_qa/en/multiple_choice_a_and_q_en/results.json
rename to evaluation_bloomz-1b7/evaluation_val/head_qa/en/multiple_choice_a_and_q_en/results.json
diff --git a/evaluation_val/head_qa/en/multiple_choice_a_and_q_with_context_en/results.json b/evaluation_bloomz-1b7/evaluation_val/head_qa/en/multiple_choice_a_and_q_with_context_en/results.json
similarity index 100%
rename from evaluation_val/head_qa/en/multiple_choice_a_and_q_with_context_en/results.json
rename to evaluation_bloomz-1b7/evaluation_val/head_qa/en/multiple_choice_a_and_q_with_context_en/results.json
diff --git a/evaluation_val/head_qa/en/multiple_choice_q_and_a_en/results.json b/evaluation_bloomz-1b7/evaluation_val/head_qa/en/multiple_choice_q_and_a_en/results.json
similarity index 100%
rename from evaluation_val/head_qa/en/multiple_choice_q_and_a_en/results.json
rename to evaluation_bloomz-1b7/evaluation_val/head_qa/en/multiple_choice_q_and_a_en/results.json
diff --git a/evaluation_val/head_qa/en/multiple_choice_q_and_a_index_en/results.json b/evaluation_bloomz-1b7/evaluation_val/head_qa/en/multiple_choice_q_and_a_index_en/results.json
similarity index 100%
rename from evaluation_val/head_qa/en/multiple_choice_q_and_a_index_en/results.json
rename to evaluation_bloomz-1b7/evaluation_val/head_qa/en/multiple_choice_q_and_a_index_en/results.json
diff --git a/evaluation_val/head_qa/en/multiple_choice_q_and_a_index_with_context_en/results.json b/evaluation_bloomz-1b7/evaluation_val/head_qa/en/multiple_choice_q_and_a_index_with_context_en/results.json
similarity index 100%
rename from evaluation_val/head_qa/en/multiple_choice_q_and_a_index_with_context_en/results.json
rename to evaluation_bloomz-1b7/evaluation_val/head_qa/en/multiple_choice_q_and_a_index_with_context_en/results.json
diff --git a/evaluation_val/head_qa/es/multiple_choice_a_and_q_en/results.json b/evaluation_bloomz-1b7/evaluation_val/head_qa/es/multiple_choice_a_and_q_en/results.json
similarity index 100%
rename from evaluation_val/head_qa/es/multiple_choice_a_and_q_en/results.json
rename to evaluation_bloomz-1b7/evaluation_val/head_qa/es/multiple_choice_a_and_q_en/results.json
diff --git a/evaluation_val/head_qa/es/multiple_choice_a_and_q_with_context_en/results.json b/evaluation_bloomz-1b7/evaluation_val/head_qa/es/multiple_choice_a_and_q_with_context_en/results.json
similarity index 100%
rename from evaluation_val/head_qa/es/multiple_choice_a_and_q_with_context_en/results.json
rename to evaluation_bloomz-1b7/evaluation_val/head_qa/es/multiple_choice_a_and_q_with_context_en/results.json
diff --git a/evaluation_val/head_qa/es/multiple_choice_q_and_a_en/results.json b/evaluation_bloomz-1b7/evaluation_val/head_qa/es/multiple_choice_q_and_a_en/results.json
similarity index 100%
rename from evaluation_val/head_qa/es/multiple_choice_q_and_a_en/results.json
rename to evaluation_bloomz-1b7/evaluation_val/head_qa/es/multiple_choice_q_and_a_en/results.json
diff --git a/evaluation_val/head_qa/es/multiple_choice_q_and_a_index_en/results.json b/evaluation_bloomz-1b7/evaluation_val/head_qa/es/multiple_choice_q_and_a_index_en/results.json
similarity index 100%
rename from evaluation_val/head_qa/es/multiple_choice_q_and_a_index_en/results.json
rename to evaluation_bloomz-1b7/evaluation_val/head_qa/es/multiple_choice_q_and_a_index_en/results.json
diff --git a/evaluation_val/head_qa/es/multiple_choice_q_and_a_index_with_context_en/results.json b/evaluation_bloomz-1b7/evaluation_val/head_qa/es/multiple_choice_q_and_a_index_with_context_en/results.json
similarity index 100%
rename from evaluation_val/head_qa/es/multiple_choice_q_and_a_index_with_context_en/results.json
rename to evaluation_bloomz-1b7/evaluation_val/head_qa/es/multiple_choice_q_and_a_index_with_context_en/results.json
diff --git a/evaluation_val/health_fact/claim_explanation_classification/results.json b/evaluation_bloomz-1b7/evaluation_val/health_fact/claim_explanation_classification/results.json
similarity index 100%
rename from evaluation_val/health_fact/claim_explanation_classification/results.json
rename to evaluation_bloomz-1b7/evaluation_val/health_fact/claim_explanation_classification/results.json
diff --git a/evaluation_val/health_fact/claim_veracity_classification_after_reading_I_believe/results.json b/evaluation_bloomz-1b7/evaluation_val/health_fact/claim_veracity_classification_after_reading_I_believe/results.json
similarity index 100%
rename from evaluation_val/health_fact/claim_veracity_classification_after_reading_I_believe/results.json
rename to evaluation_bloomz-1b7/evaluation_val/health_fact/claim_veracity_classification_after_reading_I_believe/results.json
diff --git a/evaluation_val/health_fact/claim_veracity_classification_tell_me/results.json b/evaluation_bloomz-1b7/evaluation_val/health_fact/claim_veracity_classification_tell_me/results.json
similarity index 100%
rename from evaluation_val/health_fact/claim_veracity_classification_tell_me/results.json
rename to evaluation_bloomz-1b7/evaluation_val/health_fact/claim_veracity_classification_tell_me/results.json
diff --git a/evaluation_val/hlgd/is_same_event_editor_asks/results.json b/evaluation_bloomz-1b7/evaluation_val/hlgd/is_same_event_editor_asks/results.json
similarity index 100%
rename from evaluation_val/hlgd/is_same_event_editor_asks/results.json
rename to evaluation_bloomz-1b7/evaluation_val/hlgd/is_same_event_editor_asks/results.json
diff --git a/evaluation_val/hlgd/is_same_event_interrogative_talk/results.json b/evaluation_bloomz-1b7/evaluation_val/hlgd/is_same_event_interrogative_talk/results.json
similarity index 100%
rename from evaluation_val/hlgd/is_same_event_interrogative_talk/results.json
rename to evaluation_bloomz-1b7/evaluation_val/hlgd/is_same_event_interrogative_talk/results.json
diff --git a/evaluation_val/hlgd/is_same_event_refer/results.json b/evaluation_bloomz-1b7/evaluation_val/hlgd/is_same_event_refer/results.json
similarity index 100%
rename from evaluation_val/hlgd/is_same_event_refer/results.json
rename to evaluation_bloomz-1b7/evaluation_val/hlgd/is_same_event_refer/results.json
diff --git a/evaluation_val/hlgd/is_same_event_with_time_interrogative_related/results.json b/evaluation_bloomz-1b7/evaluation_val/hlgd/is_same_event_with_time_interrogative_related/results.json
similarity index 100%
rename from evaluation_val/hlgd/is_same_event_with_time_interrogative_related/results.json
rename to evaluation_bloomz-1b7/evaluation_val/hlgd/is_same_event_with_time_interrogative_related/results.json
diff --git a/evaluation_val/hlgd/is_same_event_with_time_interrogative_talk/results.json b/evaluation_bloomz-1b7/evaluation_val/hlgd/is_same_event_with_time_interrogative_talk/results.json
similarity index 100%
rename from evaluation_val/hlgd/is_same_event_with_time_interrogative_talk/results.json
rename to evaluation_bloomz-1b7/evaluation_val/hlgd/is_same_event_with_time_interrogative_talk/results.json
diff --git a/evaluation_val/hyperpartisan_news_detection/byarticle/consider_does_it_follow_a_hyperpartisan_argumentation/results.json b/evaluation_bloomz-1b7/evaluation_val/hyperpartisan_news_detection/byarticle/consider_does_it_follow_a_hyperpartisan_argumentation/results.json
similarity index 100%
rename from evaluation_val/hyperpartisan_news_detection/byarticle/consider_does_it_follow_a_hyperpartisan_argumentation/results.json
rename to evaluation_bloomz-1b7/evaluation_val/hyperpartisan_news_detection/byarticle/consider_does_it_follow_a_hyperpartisan_argumentation/results.json
diff --git a/evaluation_val/hyperpartisan_news_detection/byarticle/consider_it_exhibits_extreme_one_sidedness/results.json b/evaluation_bloomz-1b7/evaluation_val/hyperpartisan_news_detection/byarticle/consider_it_exhibits_extreme_one_sidedness/results.json
similarity index 100%
rename from evaluation_val/hyperpartisan_news_detection/byarticle/consider_it_exhibits_extreme_one_sidedness/results.json
rename to evaluation_bloomz-1b7/evaluation_val/hyperpartisan_news_detection/byarticle/consider_it_exhibits_extreme_one_sidedness/results.json
diff --git a/evaluation_val/hyperpartisan_news_detection/byarticle/consume_with_caution/results.json b/evaluation_bloomz-1b7/evaluation_val/hyperpartisan_news_detection/byarticle/consume_with_caution/results.json
similarity index 100%
rename from evaluation_val/hyperpartisan_news_detection/byarticle/consume_with_caution/results.json
rename to evaluation_bloomz-1b7/evaluation_val/hyperpartisan_news_detection/byarticle/consume_with_caution/results.json
diff --git a/evaluation_val/hyperpartisan_news_detection/byarticle/extreme_left_wing_or_right_wing/results.json b/evaluation_bloomz-1b7/evaluation_val/hyperpartisan_news_detection/byarticle/extreme_left_wing_or_right_wing/results.json
similarity index 100%
rename from evaluation_val/hyperpartisan_news_detection/byarticle/extreme_left_wing_or_right_wing/results.json
rename to evaluation_bloomz-1b7/evaluation_val/hyperpartisan_news_detection/byarticle/extreme_left_wing_or_right_wing/results.json
diff --git a/evaluation_val/hyperpartisan_news_detection/byarticle/follows_hyperpartisan_argumentation/results.json b/evaluation_bloomz-1b7/evaluation_val/hyperpartisan_news_detection/byarticle/follows_hyperpartisan_argumentation/results.json
similarity index 100%
rename from evaluation_val/hyperpartisan_news_detection/byarticle/follows_hyperpartisan_argumentation/results.json
rename to evaluation_bloomz-1b7/evaluation_val/hyperpartisan_news_detection/byarticle/follows_hyperpartisan_argumentation/results.json
diff --git a/evaluation_val/liar/Given_statement_guess_category/results.json b/evaluation_bloomz-1b7/evaluation_val/liar/Given_statement_guess_category/results.json
similarity index 100%
rename from evaluation_val/liar/Given_statement_guess_category/results.json
rename to evaluation_bloomz-1b7/evaluation_val/liar/Given_statement_guess_category/results.json
diff --git a/evaluation_val/lince/sa_spaeng/express_sentiment/results.json b/evaluation_bloomz-1b7/evaluation_val/lince/sa_spaeng/express_sentiment/results.json
similarity index 100%
rename from evaluation_val/lince/sa_spaeng/express_sentiment/results.json
rename to evaluation_bloomz-1b7/evaluation_val/lince/sa_spaeng/express_sentiment/results.json
diff --git a/evaluation_val/lince/sa_spaeng/negation_template/results.json b/evaluation_bloomz-1b7/evaluation_val/lince/sa_spaeng/negation_template/results.json
similarity index 100%
rename from evaluation_val/lince/sa_spaeng/negation_template/results.json
rename to evaluation_bloomz-1b7/evaluation_val/lince/sa_spaeng/negation_template/results.json
diff --git a/evaluation_val/lince/sa_spaeng/original_poster_expressed_sentiment/results.json b/evaluation_bloomz-1b7/evaluation_val/lince/sa_spaeng/original_poster_expressed_sentiment/results.json
similarity index 100%
rename from evaluation_val/lince/sa_spaeng/original_poster_expressed_sentiment/results.json
rename to evaluation_bloomz-1b7/evaluation_val/lince/sa_spaeng/original_poster_expressed_sentiment/results.json
diff --git a/evaluation_val/lince/sa_spaeng/sentiment_trying_to_express/results.json b/evaluation_bloomz-1b7/evaluation_val/lince/sa_spaeng/sentiment_trying_to_express/results.json
similarity index 100%
rename from evaluation_val/lince/sa_spaeng/sentiment_trying_to_express/results.json
rename to evaluation_bloomz-1b7/evaluation_val/lince/sa_spaeng/sentiment_trying_to_express/results.json
diff --git a/evaluation_val/lince/sa_spaeng/the_author_seem/results.json b/evaluation_bloomz-1b7/evaluation_val/lince/sa_spaeng/the_author_seem/results.json
similarity index 100%
rename from evaluation_val/lince/sa_spaeng/the_author_seem/results.json
rename to evaluation_bloomz-1b7/evaluation_val/lince/sa_spaeng/the_author_seem/results.json
diff --git a/evaluation_val/math_qa/choose_correct_og/results.json b/evaluation_bloomz-1b7/evaluation_val/math_qa/choose_correct_og/results.json
similarity index 100%
rename from evaluation_val/math_qa/choose_correct_og/results.json
rename to evaluation_bloomz-1b7/evaluation_val/math_qa/choose_correct_og/results.json
diff --git a/evaluation_val/math_qa/first_choice_then_problem/results.json b/evaluation_bloomz-1b7/evaluation_val/math_qa/first_choice_then_problem/results.json
similarity index 100%
rename from evaluation_val/math_qa/first_choice_then_problem/results.json
rename to evaluation_bloomz-1b7/evaluation_val/math_qa/first_choice_then_problem/results.json
diff --git a/evaluation_val/math_qa/gre_problem/results.json b/evaluation_bloomz-1b7/evaluation_val/math_qa/gre_problem/results.json
similarity index 100%
rename from evaluation_val/math_qa/gre_problem/results.json
rename to evaluation_bloomz-1b7/evaluation_val/math_qa/gre_problem/results.json
diff --git a/evaluation_val/math_qa/pick_the_correct/results.json b/evaluation_bloomz-1b7/evaluation_val/math_qa/pick_the_correct/results.json
similarity index 100%
rename from evaluation_val/math_qa/pick_the_correct/results.json
rename to evaluation_bloomz-1b7/evaluation_val/math_qa/pick_the_correct/results.json
diff --git a/evaluation_val/math_qa/problem_set_type/results.json b/evaluation_bloomz-1b7/evaluation_val/math_qa/problem_set_type/results.json
similarity index 100%
rename from evaluation_val/math_qa/problem_set_type/results.json
rename to evaluation_bloomz-1b7/evaluation_val/math_qa/problem_set_type/results.json
diff --git a/evaluation_val/merged.csv b/evaluation_bloomz-1b7/evaluation_val/merged.csv
similarity index 100%
rename from evaluation_val/merged.csv
rename to evaluation_bloomz-1b7/evaluation_val/merged.csv
diff --git a/evaluation_val/merged.json b/evaluation_bloomz-1b7/evaluation_val/merged.json
similarity index 100%
rename from evaluation_val/merged.json
rename to evaluation_bloomz-1b7/evaluation_val/merged.json
diff --git a/evaluation_val/movie_rationales/Evidences_+_review/results.json b/evaluation_bloomz-1b7/evaluation_val/movie_rationales/Evidences_+_review/results.json
similarity index 100%
rename from evaluation_val/movie_rationales/Evidences_+_review/results.json
rename to evaluation_bloomz-1b7/evaluation_val/movie_rationales/Evidences_+_review/results.json
diff --git a/evaluation_val/movie_rationales/Evidences_sentiment_classification/results.json b/evaluation_bloomz-1b7/evaluation_val/movie_rationales/Evidences_sentiment_classification/results.json
similarity index 100%
rename from evaluation_val/movie_rationales/Evidences_sentiment_classification/results.json
rename to evaluation_bloomz-1b7/evaluation_val/movie_rationales/Evidences_sentiment_classification/results.json
diff --git a/evaluation_val/movie_rationales/Standard_binary_sentiment_analysis/results.json b/evaluation_bloomz-1b7/evaluation_val/movie_rationales/Standard_binary_sentiment_analysis/results.json
similarity index 100%
rename from evaluation_val/movie_rationales/Standard_binary_sentiment_analysis/results.json
rename to evaluation_bloomz-1b7/evaluation_val/movie_rationales/Standard_binary_sentiment_analysis/results.json
diff --git a/evaluation_val/mwsc/in-the-sentence-question-first/results.json b/evaluation_bloomz-1b7/evaluation_val/mwsc/in-the-sentence-question-first/results.json
similarity index 100%
rename from evaluation_val/mwsc/in-the-sentence-question-first/results.json
rename to evaluation_bloomz-1b7/evaluation_val/mwsc/in-the-sentence-question-first/results.json
diff --git a/evaluation_val/mwsc/in-the-sentence/results.json b/evaluation_bloomz-1b7/evaluation_val/mwsc/in-the-sentence/results.json
similarity index 100%
rename from evaluation_val/mwsc/in-the-sentence/results.json
rename to evaluation_bloomz-1b7/evaluation_val/mwsc/in-the-sentence/results.json
diff --git a/evaluation_val/mwsc/is-correct/results.json b/evaluation_bloomz-1b7/evaluation_val/mwsc/is-correct/results.json
similarity index 100%
rename from evaluation_val/mwsc/is-correct/results.json
rename to evaluation_bloomz-1b7/evaluation_val/mwsc/is-correct/results.json
diff --git a/evaluation_val/mwsc/options-or/results.json b/evaluation_bloomz-1b7/evaluation_val/mwsc/options-or/results.json
similarity index 100%
rename from evaluation_val/mwsc/options-or/results.json
rename to evaluation_bloomz-1b7/evaluation_val/mwsc/options-or/results.json
diff --git a/evaluation_val/mwsc/what-think/results.json b/evaluation_bloomz-1b7/evaluation_val/mwsc/what-think/results.json
similarity index 100%
rename from evaluation_val/mwsc/what-think/results.json
rename to evaluation_bloomz-1b7/evaluation_val/mwsc/what-think/results.json
diff --git a/evaluation_val/onestop_english/ara_context/results.json b/evaluation_bloomz-1b7/evaluation_val/onestop_english/ara_context/results.json
similarity index 100%
rename from evaluation_val/onestop_english/ara_context/results.json
rename to evaluation_bloomz-1b7/evaluation_val/onestop_english/ara_context/results.json
diff --git a/evaluation_val/onestop_english/assess/results.json b/evaluation_bloomz-1b7/evaluation_val/onestop_english/assess/results.json
similarity index 100%
rename from evaluation_val/onestop_english/assess/results.json
rename to evaluation_bloomz-1b7/evaluation_val/onestop_english/assess/results.json
diff --git a/evaluation_val/onestop_english/determine_reading_level_from_the_first_three_sentences/results.json b/evaluation_bloomz-1b7/evaluation_val/onestop_english/determine_reading_level_from_the_first_three_sentences/results.json
similarity index 100%
rename from evaluation_val/onestop_english/determine_reading_level_from_the_first_three_sentences/results.json
rename to evaluation_bloomz-1b7/evaluation_val/onestop_english/determine_reading_level_from_the_first_three_sentences/results.json
diff --git a/evaluation_val/onestop_english/esl_context/results.json b/evaluation_bloomz-1b7/evaluation_val/onestop_english/esl_context/results.json
similarity index 100%
rename from evaluation_val/onestop_english/esl_context/results.json
rename to evaluation_bloomz-1b7/evaluation_val/onestop_english/esl_context/results.json
diff --git a/evaluation_val/onestop_english/esl_variation/results.json b/evaluation_bloomz-1b7/evaluation_val/onestop_english/esl_variation/results.json
similarity index 100%
rename from evaluation_val/onestop_english/esl_variation/results.json
rename to evaluation_bloomz-1b7/evaluation_val/onestop_english/esl_variation/results.json
diff --git a/evaluation_val/poem_sentiment/guess_sentiment_without_options_variation_1/results.json b/evaluation_bloomz-1b7/evaluation_val/poem_sentiment/guess_sentiment_without_options_variation_1/results.json
similarity index 100%
rename from evaluation_val/poem_sentiment/guess_sentiment_without_options_variation_1/results.json
rename to evaluation_bloomz-1b7/evaluation_val/poem_sentiment/guess_sentiment_without_options_variation_1/results.json
diff --git a/evaluation_val/poem_sentiment/most_appropriate_sentiment/results.json b/evaluation_bloomz-1b7/evaluation_val/poem_sentiment/most_appropriate_sentiment/results.json
similarity index 100%
rename from evaluation_val/poem_sentiment/most_appropriate_sentiment/results.json
rename to evaluation_bloomz-1b7/evaluation_val/poem_sentiment/most_appropriate_sentiment/results.json
diff --git a/evaluation_val/poem_sentiment/positive_or_negative_sentiment_variation_1/results.json b/evaluation_bloomz-1b7/evaluation_val/poem_sentiment/positive_or_negative_sentiment_variation_1/results.json
similarity index 100%
rename from evaluation_val/poem_sentiment/positive_or_negative_sentiment_variation_1/results.json
rename to evaluation_bloomz-1b7/evaluation_val/poem_sentiment/positive_or_negative_sentiment_variation_1/results.json
diff --git a/evaluation_val/poem_sentiment/positive_or_negative_sentiment_variation_2/results.json b/evaluation_bloomz-1b7/evaluation_val/poem_sentiment/positive_or_negative_sentiment_variation_2/results.json
similarity index 100%
rename from evaluation_val/poem_sentiment/positive_or_negative_sentiment_variation_2/results.json
rename to evaluation_bloomz-1b7/evaluation_val/poem_sentiment/positive_or_negative_sentiment_variation_2/results.json
diff --git a/evaluation_val/poem_sentiment/question_answer_format/results.json b/evaluation_bloomz-1b7/evaluation_val/poem_sentiment/question_answer_format/results.json
similarity index 100%
rename from evaluation_val/poem_sentiment/question_answer_format/results.json
rename to evaluation_bloomz-1b7/evaluation_val/poem_sentiment/question_answer_format/results.json
diff --git a/evaluation_val/pubmed_qa/pqa_labeled/Long_Answer_to_Final_Decision/results.json b/evaluation_bloomz-1b7/evaluation_val/pubmed_qa/pqa_labeled/Long_Answer_to_Final_Decision/results.json
similarity index 100%
rename from evaluation_val/pubmed_qa/pqa_labeled/Long_Answer_to_Final_Decision/results.json
rename to evaluation_bloomz-1b7/evaluation_val/pubmed_qa/pqa_labeled/Long_Answer_to_Final_Decision/results.json
diff --git a/evaluation_val/pubmed_qa/pqa_labeled/Question_Answering_(Short)/results.json b/evaluation_bloomz-1b7/evaluation_val/pubmed_qa/pqa_labeled/Question_Answering_(Short)/results.json
similarity index 100%
rename from evaluation_val/pubmed_qa/pqa_labeled/Question_Answering_(Short)/results.json
rename to evaluation_bloomz-1b7/evaluation_val/pubmed_qa/pqa_labeled/Question_Answering_(Short)/results.json
diff --git a/evaluation_val/riddle_sense/answer_given_question_without_options/results.json b/evaluation_bloomz-1b7/evaluation_val/riddle_sense/answer_given_question_without_options/results.json
similarity index 100%
rename from evaluation_val/riddle_sense/answer_given_question_without_options/results.json
rename to evaluation_bloomz-1b7/evaluation_val/riddle_sense/answer_given_question_without_options/results.json
diff --git a/evaluation_val/riddle_sense/most_suitable_answer/results.json b/evaluation_bloomz-1b7/evaluation_val/riddle_sense/most_suitable_answer/results.json
similarity index 100%
rename from evaluation_val/riddle_sense/most_suitable_answer/results.json
rename to evaluation_bloomz-1b7/evaluation_val/riddle_sense/most_suitable_answer/results.json
diff --git a/evaluation_val/riddle_sense/question_answering/results.json b/evaluation_bloomz-1b7/evaluation_val/riddle_sense/question_answering/results.json
similarity index 100%
rename from evaluation_val/riddle_sense/question_answering/results.json
rename to evaluation_bloomz-1b7/evaluation_val/riddle_sense/question_answering/results.json
diff --git a/evaluation_val/riddle_sense/question_to_answer_index/results.json b/evaluation_bloomz-1b7/evaluation_val/riddle_sense/question_to_answer_index/results.json
similarity index 100%
rename from evaluation_val/riddle_sense/question_to_answer_index/results.json
rename to evaluation_bloomz-1b7/evaluation_val/riddle_sense/question_to_answer_index/results.json
diff --git a/evaluation_val/scicite/Classify_intent/results.json b/evaluation_bloomz-1b7/evaluation_val/scicite/Classify_intent/results.json
similarity index 100%
rename from evaluation_val/scicite/Classify_intent/results.json
rename to evaluation_bloomz-1b7/evaluation_val/scicite/Classify_intent/results.json
diff --git a/evaluation_val/scicite/Classify_intent_(choices_first)/results.json b/evaluation_bloomz-1b7/evaluation_val/scicite/Classify_intent_(choices_first)/results.json
similarity index 100%
rename from evaluation_val/scicite/Classify_intent_(choices_first)/results.json
rename to evaluation_bloomz-1b7/evaluation_val/scicite/Classify_intent_(choices_first)/results.json
diff --git a/evaluation_val/scicite/Classify_intent_(select_choice)/results.json b/evaluation_bloomz-1b7/evaluation_val/scicite/Classify_intent_(select_choice)/results.json
similarity index 100%
rename from evaluation_val/scicite/Classify_intent_(select_choice)/results.json
rename to evaluation_bloomz-1b7/evaluation_val/scicite/Classify_intent_(select_choice)/results.json
diff --git a/evaluation_val/scicite/Classify_intent_w_section_(select_choice)/results.json b/evaluation_bloomz-1b7/evaluation_val/scicite/Classify_intent_w_section_(select_choice)/results.json
similarity index 100%
rename from evaluation_val/scicite/Classify_intent_w_section_(select_choice)/results.json
rename to evaluation_bloomz-1b7/evaluation_val/scicite/Classify_intent_w_section_(select_choice)/results.json
diff --git a/evaluation_val/scicite/can_describe/results.json b/evaluation_bloomz-1b7/evaluation_val/scicite/can_describe/results.json
similarity index 100%
rename from evaluation_val/scicite/can_describe/results.json
rename to evaluation_bloomz-1b7/evaluation_val/scicite/can_describe/results.json
diff --git a/evaluation_val/selqa/answer_selection_analysis/is-he-talking-about/results.json b/evaluation_bloomz-1b7/evaluation_val/selqa/answer_selection_analysis/is-he-talking-about/results.json
similarity index 100%
rename from evaluation_val/selqa/answer_selection_analysis/is-he-talking-about/results.json
rename to evaluation_bloomz-1b7/evaluation_val/selqa/answer_selection_analysis/is-he-talking-about/results.json
diff --git a/evaluation_val/selqa/answer_selection_analysis/make-sense-rand/results.json b/evaluation_bloomz-1b7/evaluation_val/selqa/answer_selection_analysis/make-sense-rand/results.json
similarity index 100%
rename from evaluation_val/selqa/answer_selection_analysis/make-sense-rand/results.json
rename to evaluation_bloomz-1b7/evaluation_val/selqa/answer_selection_analysis/make-sense-rand/results.json
diff --git a/evaluation_val/selqa/answer_selection_analysis/which-answer-1st-vs-random/results.json b/evaluation_bloomz-1b7/evaluation_val/selqa/answer_selection_analysis/which-answer-1st-vs-random/results.json
similarity index 100%
rename from evaluation_val/selqa/answer_selection_analysis/which-answer-1st-vs-random/results.json
rename to evaluation_bloomz-1b7/evaluation_val/selqa/answer_selection_analysis/which-answer-1st-vs-random/results.json
diff --git a/evaluation_val/selqa/answer_selection_analysis/would-make-sense-qu-rand/results.json b/evaluation_bloomz-1b7/evaluation_val/selqa/answer_selection_analysis/would-make-sense-qu-rand/results.json
similarity index 100%
rename from evaluation_val/selqa/answer_selection_analysis/would-make-sense-qu-rand/results.json
rename to evaluation_bloomz-1b7/evaluation_val/selqa/answer_selection_analysis/would-make-sense-qu-rand/results.json
diff --git a/evaluation_val/snips_built_in_intents/categorize_query/results.json b/evaluation_bloomz-1b7/evaluation_val/snips_built_in_intents/categorize_query/results.json
similarity index 100%
rename from evaluation_val/snips_built_in_intents/categorize_query/results.json
rename to evaluation_bloomz-1b7/evaluation_val/snips_built_in_intents/categorize_query/results.json
diff --git a/evaluation_val/snips_built_in_intents/categorize_query_brief/results.json b/evaluation_bloomz-1b7/evaluation_val/snips_built_in_intents/categorize_query_brief/results.json
similarity index 100%
rename from evaluation_val/snips_built_in_intents/categorize_query_brief/results.json
rename to evaluation_bloomz-1b7/evaluation_val/snips_built_in_intents/categorize_query_brief/results.json
diff --git a/evaluation_val/snips_built_in_intents/intent_query/results.json b/evaluation_bloomz-1b7/evaluation_val/snips_built_in_intents/intent_query/results.json
similarity index 100%
rename from evaluation_val/snips_built_in_intents/intent_query/results.json
rename to evaluation_bloomz-1b7/evaluation_val/snips_built_in_intents/intent_query/results.json
diff --git a/evaluation_val/snips_built_in_intents/query_intent/results.json b/evaluation_bloomz-1b7/evaluation_val/snips_built_in_intents/query_intent/results.json
similarity index 100%
rename from evaluation_val/snips_built_in_intents/query_intent/results.json
rename to evaluation_bloomz-1b7/evaluation_val/snips_built_in_intents/query_intent/results.json
diff --git a/evaluation_val/snips_built_in_intents/voice_intent/results.json b/evaluation_bloomz-1b7/evaluation_val/snips_built_in_intents/voice_intent/results.json
similarity index 100%
rename from evaluation_val/snips_built_in_intents/voice_intent/results.json
rename to evaluation_bloomz-1b7/evaluation_val/snips_built_in_intents/voice_intent/results.json
diff --git a/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=layman_summ_es.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=layman_summ_es.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
similarity index 100%
rename from evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=layman_summ_es.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=layman_summ_es.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
diff --git a/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=palm_prompt.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=palm_prompt.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
similarity index 100%
rename from evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=palm_prompt.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=palm_prompt.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
diff --git a/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=summarise_this_in_es_few_sentences.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=summarise_this_in_es_few_sentences.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
similarity index 100%
rename from evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=summarise_this_in_es_few_sentences.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=summarise_this_in_es_few_sentences.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
diff --git a/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-en-fr-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-en-fr-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
similarity index 100%
rename from evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-en-fr-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-en-fr-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
diff --git a/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
similarity index 100%
rename from evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
diff --git a/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-fr-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-fr-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json
similarity index 100%
rename from evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-fr-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-fr-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json
diff --git a/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
similarity index 100%
rename from evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
diff --git a/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=gpt3-en-fr.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=gpt3-en-fr.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
similarity index 100%
rename from evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=gpt3-en-fr.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=gpt3-en-fr.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
diff --git a/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=gpt3-fr-en.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=gpt3-fr-en.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json
similarity index 100%
rename from evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=gpt3-fr-en.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=gpt3-fr-en.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json
diff --git a/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=version-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=version-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
similarity index 100%
rename from evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=version-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=version-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
diff --git a/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=version-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=version-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
similarity index 100%
rename from evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=version-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=version-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
diff --git a/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=xglm-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=xglm-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
similarity index 100%
rename from evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=xglm-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=xglm-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
diff --git a/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=xglm-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=xglm-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json
similarity index 100%
rename from evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=xglm-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=xglm-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json
diff --git a/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-en-hi-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-en-hi-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
similarity index 100%
rename from evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-en-hi-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-en-hi-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
diff --git a/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json
similarity index 100%
rename from evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json
diff --git a/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-hi-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-hi-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.json
similarity index 100%
rename from evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-hi-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-hi-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.json
diff --git a/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.json
similarity index 100%
rename from evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.json
diff --git a/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=gpt-3-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=gpt-3-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
similarity index 100%
rename from evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=gpt-3-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=gpt-3-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
diff --git a/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=gpt-3-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=gpt-3-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
similarity index 100%
rename from evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=gpt-3-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=gpt-3-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
diff --git a/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=version-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=version-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json
similarity index 100%
rename from evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=version-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=version-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json
diff --git a/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=version-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=version-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
similarity index 100%
rename from evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=version-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=version-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
diff --git a/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=xglm-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=xglm-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
similarity index 100%
rename from evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=xglm-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=xglm-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
diff --git a/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=xglm-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=xglm-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.json
similarity index 100%
rename from evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=xglm-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/agg.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=xglm-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.json
diff --git a/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=layman_summ_es.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.jsonl b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=layman_summ_es.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.jsonl
similarity index 100%
rename from evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=layman_summ_es.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.jsonl
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=layman_summ_es.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.jsonl
diff --git a/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=palm_prompt.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.jsonl b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=palm_prompt.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.jsonl
similarity index 100%
rename from evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=palm_prompt.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.jsonl
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=palm_prompt.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.jsonl
diff --git a/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=summarise_this_in_es_few_sentences.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.jsonl b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=summarise_this_in_es_few_sentences.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.jsonl
similarity index 100%
rename from evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=summarise_this_in_es_few_sentences.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.jsonl
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=summarise_this_in_es_few_sentences.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.jsonl
diff --git a/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-en-fr-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.jsonl b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-en-fr-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.jsonl
similarity index 100%
rename from evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-en-fr-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.jsonl
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-en-fr-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.jsonl
diff --git a/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.jsonl b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.jsonl
similarity index 100%
rename from evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.jsonl
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.jsonl
diff --git a/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-fr-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.jsonl b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-fr-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.jsonl
similarity index 100%
rename from evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-fr-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.jsonl
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-fr-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.jsonl
diff --git a/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.jsonl b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.jsonl
similarity index 100%
rename from evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.jsonl
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.jsonl
diff --git a/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=gpt3-en-fr.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.jsonl b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=gpt3-en-fr.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.jsonl
similarity index 100%
rename from evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=gpt3-en-fr.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.jsonl
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=gpt3-en-fr.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.jsonl
diff --git a/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=gpt3-fr-en.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.jsonl b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=gpt3-fr-en.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.jsonl
similarity index 100%
rename from evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=gpt3-fr-en.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.jsonl
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=gpt3-fr-en.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.jsonl
diff --git a/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=version-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.jsonl b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=version-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.jsonl
similarity index 100%
rename from evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=version-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.jsonl
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=version-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.jsonl
diff --git a/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=version-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.jsonl b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=version-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.jsonl
similarity index 100%
rename from evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=version-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.jsonl
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=version-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.jsonl
diff --git a/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=xglm-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.jsonl b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=xglm-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.jsonl
similarity index 100%
rename from evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=xglm-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.jsonl
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=xglm-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.jsonl
diff --git a/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=xglm-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.jsonl b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=xglm-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.jsonl
similarity index 100%
rename from evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=xglm-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.jsonl
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=xglm-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.jsonl
diff --git a/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-en-hi-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.jsonl b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-en-hi-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.jsonl
similarity index 100%
rename from evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-en-hi-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.jsonl
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-en-hi-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.jsonl
diff --git a/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.jsonl b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.jsonl
similarity index 100%
rename from evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.jsonl
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.jsonl
diff --git a/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-hi-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.jsonl b/evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-hi-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.jsonl
similarity index 100%
rename from evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-hi-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.jsonl
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_fr_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-hi-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.jsonl
diff --git a/evaluation_val/wmt14_hi_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.jsonl b/evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.jsonl
similarity index 100%
rename from evaluation_val/wmt14_hi_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.jsonl
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.jsonl
diff --git a/evaluation_val/wmt14_hi_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=gpt-3-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.jsonl b/evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=gpt-3-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.jsonl
similarity index 100%
rename from evaluation_val/wmt14_hi_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=gpt-3-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.jsonl
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=gpt-3-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.jsonl
diff --git a/evaluation_val/wmt14_hi_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=gpt-3-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.jsonl b/evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=gpt-3-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.jsonl
similarity index 100%
rename from evaluation_val/wmt14_hi_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=gpt-3-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.jsonl
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=gpt-3-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.jsonl
diff --git a/evaluation_val/wmt14_hi_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=version-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.jsonl b/evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=version-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.jsonl
similarity index 100%
rename from evaluation_val/wmt14_hi_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=version-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.jsonl
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=version-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.jsonl
diff --git a/evaluation_val/wmt14_hi_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=version-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.jsonl b/evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=version-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.jsonl
similarity index 100%
rename from evaluation_val/wmt14_hi_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=version-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.jsonl
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=version-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.jsonl
diff --git a/evaluation_val/wmt14_hi_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=xglm-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.jsonl b/evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=xglm-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.jsonl
similarity index 100%
rename from evaluation_val/wmt14_hi_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=xglm-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.jsonl
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=xglm-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.jsonl
diff --git a/evaluation_val/wmt14_hi_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=xglm-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.jsonl b/evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=xglm-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.jsonl
similarity index 100%
rename from evaluation_val/wmt14_hi_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=xglm-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.jsonl
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/examples.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=xglm-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.jsonl
diff --git a/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=layman_summ_es.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=layman_summ_es.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
similarity index 100%
rename from evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=layman_summ_es.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=layman_summ_es.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
diff --git a/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=palm_prompt.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=palm_prompt.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
similarity index 100%
rename from evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=palm_prompt.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=palm_prompt.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
diff --git a/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=summarise_this_in_es_few_sentences.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=summarise_this_in_es_few_sentences.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
similarity index 100%
rename from evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=summarise_this_in_es_few_sentences.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=mlsum_es.templates=summarise_this_in_es_few_sentences.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
diff --git a/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-en-fr-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-en-fr-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
similarity index 100%
rename from evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-en-fr-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-en-fr-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
diff --git a/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
similarity index 100%
rename from evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
diff --git a/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-fr-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-fr-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json
similarity index 100%
rename from evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-fr-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-fr-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json
diff --git a/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
similarity index 100%
rename from evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=a_good_translation-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
diff --git a/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=gpt3-en-fr.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=gpt3-en-fr.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
similarity index 100%
rename from evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=gpt3-en-fr.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=gpt3-en-fr.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
diff --git a/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=gpt3-fr-en.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=gpt3-fr-en.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json
similarity index 100%
rename from evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=gpt3-fr-en.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=gpt3-fr-en.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json
diff --git a/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=version-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=version-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
similarity index 100%
rename from evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=version-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=version-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
diff --git a/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=version-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=version-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
similarity index 100%
rename from evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=version-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=version-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
diff --git a/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=xglm-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=xglm-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
similarity index 100%
rename from evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=xglm-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=xglm-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:36.json
diff --git a/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=xglm-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=xglm-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json
similarity index 100%
rename from evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=xglm-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_fr_en.templates=xglm-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json
diff --git a/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-en-hi-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-en-hi-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
similarity index 100%
rename from evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-en-hi-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-en-hi-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
diff --git a/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json
similarity index 100%
rename from evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json
diff --git a/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-hi-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-hi-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.json
similarity index 100%
rename from evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-hi-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-hi-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.json
diff --git a/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.json
similarity index 100%
rename from evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=a_good_translation-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.json
diff --git a/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=gpt-3-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=gpt-3-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
similarity index 100%
rename from evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=gpt-3-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=gpt-3-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
diff --git a/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=gpt-3-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=gpt-3-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
similarity index 100%
rename from evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=gpt-3-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=gpt-3-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
diff --git a/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=version-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=version-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json
similarity index 100%
rename from evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=version-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=version-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:48.json
diff --git a/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=version-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=version-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
similarity index 100%
rename from evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=version-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=version-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
diff --git a/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=xglm-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=xglm-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
similarity index 100%
rename from evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=xglm-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=xglm-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:46.json
diff --git a/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=xglm-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.json b/evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=xglm-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.json
similarity index 100%
rename from evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=xglm-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.json
rename to evaluation_bloomz-1b7/evaluation_val/wmt14_hi_en/slim.limited=3000.model=1b3t0_xp3capmixlossseq_global_step2000.task=wmt14_hi_en.templates=xglm-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-10-09T08:54:47.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/ar/GPT-3_style_arht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ar/GPT-3_style_arht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/ar/GPT-3_style_arht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ar/GPT-3_style_arht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/ar/GPT-3_style_armt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ar/GPT-3_style_armt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/ar/GPT-3_style_armt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ar/GPT-3_style_armt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/ar/MNLI_crowdsource_arht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ar/MNLI_crowdsource_arht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/ar/MNLI_crowdsource_arht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ar/MNLI_crowdsource_arht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/ar/MNLI_crowdsource_armt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ar/MNLI_crowdsource_armt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/ar/MNLI_crowdsource_armt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ar/MNLI_crowdsource_armt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/ar/can_we_infer_arht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ar/can_we_infer_arht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/ar/can_we_infer_arht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ar/can_we_infer_arht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/ar/can_we_infer_armt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ar/can_we_infer_armt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/ar/can_we_infer_armt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ar/can_we_infer_armt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/ar/guaranteed_possible_impossible_arht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ar/guaranteed_possible_impossible_arht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/ar/guaranteed_possible_impossible_arht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ar/guaranteed_possible_impossible_arht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/ar/guaranteed_possible_impossible_armt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ar/guaranteed_possible_impossible_armt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/ar/guaranteed_possible_impossible_armt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ar/guaranteed_possible_impossible_armt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/ar/justified_in_saying_arht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ar/justified_in_saying_arht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/ar/justified_in_saying_arht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ar/justified_in_saying_arht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/ar/justified_in_saying_armt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ar/justified_in_saying_armt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/ar/justified_in_saying_armt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ar/justified_in_saying_armt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/es/GPT-3_style_esht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/es/GPT-3_style_esht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/es/GPT-3_style_esht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/es/GPT-3_style_esht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/es/GPT-3_style_esmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/es/GPT-3_style_esmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/es/GPT-3_style_esmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/es/GPT-3_style_esmt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/es/MNLI_crowdsource_esht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/es/MNLI_crowdsource_esht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/es/MNLI_crowdsource_esht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/es/MNLI_crowdsource_esht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/es/MNLI_crowdsource_esmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/es/MNLI_crowdsource_esmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/es/MNLI_crowdsource_esmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/es/MNLI_crowdsource_esmt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/es/can_we_infer_esht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/es/can_we_infer_esht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/es/can_we_infer_esht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/es/can_we_infer_esht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/es/can_we_infer_esmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/es/can_we_infer_esmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/es/can_we_infer_esmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/es/can_we_infer_esmt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/es/guaranteed_possible_impossible_esht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/es/guaranteed_possible_impossible_esht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/es/guaranteed_possible_impossible_esht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/es/guaranteed_possible_impossible_esht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/es/guaranteed_possible_impossible_esmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/es/guaranteed_possible_impossible_esmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/es/guaranteed_possible_impossible_esmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/es/guaranteed_possible_impossible_esmt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/es/justified_in_saying_esht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/es/justified_in_saying_esht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/es/justified_in_saying_esht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/es/justified_in_saying_esht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/es/justified_in_saying_esmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/es/justified_in_saying_esmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/es/justified_in_saying_esmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/es/justified_in_saying_esmt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/fr/GPT-3_style_frht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/fr/GPT-3_style_frht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/fr/GPT-3_style_frht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/fr/GPT-3_style_frht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/fr/GPT-3_style_frmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/fr/GPT-3_style_frmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/fr/GPT-3_style_frmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/fr/GPT-3_style_frmt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/fr/MNLI_crowdsource_frht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/fr/MNLI_crowdsource_frht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/fr/MNLI_crowdsource_frht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/fr/MNLI_crowdsource_frht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/fr/MNLI_crowdsource_frmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/fr/MNLI_crowdsource_frmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/fr/MNLI_crowdsource_frmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/fr/MNLI_crowdsource_frmt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/fr/can_we_infer_frht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/fr/can_we_infer_frht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/fr/can_we_infer_frht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/fr/can_we_infer_frht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/fr/can_we_infer_frmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/fr/can_we_infer_frmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/fr/can_we_infer_frmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/fr/can_we_infer_frmt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/fr/guaranteed_possible_impossible_frht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/fr/guaranteed_possible_impossible_frht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/fr/guaranteed_possible_impossible_frht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/fr/guaranteed_possible_impossible_frht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/fr/guaranteed_possible_impossible_frmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/fr/guaranteed_possible_impossible_frmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/fr/guaranteed_possible_impossible_frmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/fr/guaranteed_possible_impossible_frmt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/fr/justified_in_saying_frht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/fr/justified_in_saying_frht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/fr/justified_in_saying_frht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/fr/justified_in_saying_frht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/fr/justified_in_saying_frmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/fr/justified_in_saying_frmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/fr/justified_in_saying_frmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/fr/justified_in_saying_frmt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/hi/GPT-3_style_hiht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/hi/GPT-3_style_hiht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/hi/GPT-3_style_hiht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/hi/GPT-3_style_hiht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/hi/GPT-3_style_himt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/hi/GPT-3_style_himt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/hi/GPT-3_style_himt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/hi/GPT-3_style_himt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/hi/MNLI_crowdsource_hiht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/hi/MNLI_crowdsource_hiht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/hi/MNLI_crowdsource_hiht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/hi/MNLI_crowdsource_hiht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/hi/MNLI_crowdsource_himt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/hi/MNLI_crowdsource_himt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/hi/MNLI_crowdsource_himt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/hi/MNLI_crowdsource_himt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/hi/can_we_infer_hiht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/hi/can_we_infer_hiht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/hi/can_we_infer_hiht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/hi/can_we_infer_hiht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/hi/can_we_infer_himt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/hi/can_we_infer_himt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/hi/can_we_infer_himt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/hi/can_we_infer_himt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/hi/guaranteed_possible_impossible_hiht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/hi/guaranteed_possible_impossible_hiht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/hi/guaranteed_possible_impossible_hiht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/hi/guaranteed_possible_impossible_hiht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/hi/guaranteed_possible_impossible_himt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/hi/guaranteed_possible_impossible_himt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/hi/guaranteed_possible_impossible_himt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/hi/guaranteed_possible_impossible_himt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/hi/justified_in_saying_hiht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/hi/justified_in_saying_hiht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/hi/justified_in_saying_hiht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/hi/justified_in_saying_hiht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/hi/justified_in_saying_himt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/hi/justified_in_saying_himt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/hi/justified_in_saying_himt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/hi/justified_in_saying_himt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/sw/GPT-3_style_swht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/sw/GPT-3_style_swht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/sw/GPT-3_style_swht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/sw/GPT-3_style_swht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/sw/GPT-3_style_swmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/sw/GPT-3_style_swmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/sw/GPT-3_style_swmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/sw/GPT-3_style_swmt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/sw/MNLI_crowdsource_swht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/sw/MNLI_crowdsource_swht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/sw/MNLI_crowdsource_swht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/sw/MNLI_crowdsource_swht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/sw/MNLI_crowdsource_swmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/sw/MNLI_crowdsource_swmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/sw/MNLI_crowdsource_swmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/sw/MNLI_crowdsource_swmt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/sw/can_we_infer_swht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/sw/can_we_infer_swht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/sw/can_we_infer_swht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/sw/can_we_infer_swht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/sw/can_we_infer_swmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/sw/can_we_infer_swmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/sw/can_we_infer_swmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/sw/can_we_infer_swmt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/sw/guaranteed_possible_impossible_swht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/sw/guaranteed_possible_impossible_swht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/sw/guaranteed_possible_impossible_swht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/sw/guaranteed_possible_impossible_swht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/sw/guaranteed_possible_impossible_swmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/sw/guaranteed_possible_impossible_swmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/sw/guaranteed_possible_impossible_swmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/sw/guaranteed_possible_impossible_swmt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/sw/justified_in_saying_swht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/sw/justified_in_saying_swht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/sw/justified_in_saying_swht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/sw/justified_in_saying_swht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/sw/justified_in_saying_swmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/sw/justified_in_saying_swmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/sw/justified_in_saying_swmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/sw/justified_in_saying_swmt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/ur/GPT-3_style_urht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ur/GPT-3_style_urht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/ur/GPT-3_style_urht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ur/GPT-3_style_urht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/ur/GPT-3_style_urmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ur/GPT-3_style_urmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/ur/GPT-3_style_urmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ur/GPT-3_style_urmt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/ur/MNLI_crowdsource_urht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ur/MNLI_crowdsource_urht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/ur/MNLI_crowdsource_urht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ur/MNLI_crowdsource_urht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/ur/MNLI_crowdsource_urmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ur/MNLI_crowdsource_urmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/ur/MNLI_crowdsource_urmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ur/MNLI_crowdsource_urmt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/ur/can_we_infer_urht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ur/can_we_infer_urht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/ur/can_we_infer_urht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ur/can_we_infer_urht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/ur/can_we_infer_urmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ur/can_we_infer_urmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/ur/can_we_infer_urmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ur/can_we_infer_urmt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/ur/guaranteed_possible_impossible_urht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ur/guaranteed_possible_impossible_urht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/ur/guaranteed_possible_impossible_urht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ur/guaranteed_possible_impossible_urht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/ur/guaranteed_possible_impossible_urmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ur/guaranteed_possible_impossible_urmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/ur/guaranteed_possible_impossible_urmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ur/guaranteed_possible_impossible_urmt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/ur/justified_in_saying_urht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ur/justified_in_saying_urht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/ur/justified_in_saying_urht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ur/justified_in_saying_urht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/ur/justified_in_saying_urmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ur/justified_in_saying_urmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/ur/justified_in_saying_urmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/ur/justified_in_saying_urmt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/vi/GPT-3_style_viht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/vi/GPT-3_style_viht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/vi/GPT-3_style_viht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/vi/GPT-3_style_viht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/vi/GPT-3_style_vimt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/vi/GPT-3_style_vimt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/vi/GPT-3_style_vimt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/vi/GPT-3_style_vimt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/vi/MNLI_crowdsource_viht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/vi/MNLI_crowdsource_viht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/vi/MNLI_crowdsource_viht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/vi/MNLI_crowdsource_viht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/vi/MNLI_crowdsource_vimt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/vi/MNLI_crowdsource_vimt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/vi/MNLI_crowdsource_vimt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/vi/MNLI_crowdsource_vimt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/vi/can_we_infer_viht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/vi/can_we_infer_viht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/vi/can_we_infer_viht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/vi/can_we_infer_viht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/vi/can_we_infer_vimt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/vi/can_we_infer_vimt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/vi/can_we_infer_vimt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/vi/can_we_infer_vimt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/vi/guaranteed_possible_impossible_viht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/vi/guaranteed_possible_impossible_viht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/vi/guaranteed_possible_impossible_viht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/vi/guaranteed_possible_impossible_viht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/vi/guaranteed_possible_impossible_vimt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/vi/guaranteed_possible_impossible_vimt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/vi/guaranteed_possible_impossible_vimt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/vi/guaranteed_possible_impossible_vimt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/vi/justified_in_saying_viht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/vi/justified_in_saying_viht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/vi/justified_in_saying_viht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/vi/justified_in_saying_viht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/vi/justified_in_saying_vimt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/vi/justified_in_saying_vimt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/vi/justified_in_saying_vimt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/vi/justified_in_saying_vimt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/zh/GPT-3_style_zhht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/zh/GPT-3_style_zhht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/zh/GPT-3_style_zhht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/zh/GPT-3_style_zhht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/zh/GPT-3_style_zhmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/zh/GPT-3_style_zhmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/zh/GPT-3_style_zhmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/zh/GPT-3_style_zhmt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/zh/MNLI_crowdsource_zhht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/zh/MNLI_crowdsource_zhht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/zh/MNLI_crowdsource_zhht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/zh/MNLI_crowdsource_zhht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/zh/MNLI_crowdsource_zhmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/zh/MNLI_crowdsource_zhmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/zh/MNLI_crowdsource_zhmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/zh/MNLI_crowdsource_zhmt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/zh/can_we_infer_zhht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/zh/can_we_infer_zhht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/zh/can_we_infer_zhht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/zh/can_we_infer_zhht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/zh/can_we_infer_zhmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/zh/can_we_infer_zhmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/zh/can_we_infer_zhmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/zh/can_we_infer_zhmt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/zh/guaranteed_possible_impossible_zhht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/zh/guaranteed_possible_impossible_zhht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/zh/guaranteed_possible_impossible_zhht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/zh/guaranteed_possible_impossible_zhht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/zh/guaranteed_possible_impossible_zhmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/zh/guaranteed_possible_impossible_zhmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/zh/guaranteed_possible_impossible_zhmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/zh/guaranteed_possible_impossible_zhmt/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/zh/justified_in_saying_zhht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/zh/justified_in_saying_zhht/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/zh/justified_in_saying_zhht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/zh/justified_in_saying_zhht/results.json
diff --git a/evaluation_xnlimtht/evaluation/xnli/zh/justified_in_saying_zhmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/zh/justified_in_saying_zhmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/evaluation/xnli/zh/justified_in_saying_zhmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/evaluation/xnli/zh/justified_in_saying_zhmt/results.json
diff --git a/evaluation_xnlimtht/xnli/ar/GPT-3_style_arht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ar/GPT-3_style_arht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/ar/GPT-3_style_arht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ar/GPT-3_style_arht/results.json
diff --git a/evaluation_xnlimtht/xnli/ar/GPT-3_style_armt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ar/GPT-3_style_armt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/ar/GPT-3_style_armt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ar/GPT-3_style_armt/results.json
diff --git a/evaluation_xnlimtht/xnli/ar/MNLI_crowdsource_arht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ar/MNLI_crowdsource_arht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/ar/MNLI_crowdsource_arht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ar/MNLI_crowdsource_arht/results.json
diff --git a/evaluation_xnlimtht/xnli/ar/MNLI_crowdsource_armt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ar/MNLI_crowdsource_armt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/ar/MNLI_crowdsource_armt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ar/MNLI_crowdsource_armt/results.json
diff --git a/evaluation_xnlimtht/xnli/ar/can_we_infer_arht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ar/can_we_infer_arht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/ar/can_we_infer_arht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ar/can_we_infer_arht/results.json
diff --git a/evaluation_xnlimtht/xnli/ar/can_we_infer_armt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ar/can_we_infer_armt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/ar/can_we_infer_armt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ar/can_we_infer_armt/results.json
diff --git a/evaluation_xnlimtht/xnli/ar/guaranteed_possible_impossible_arht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ar/guaranteed_possible_impossible_arht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/ar/guaranteed_possible_impossible_arht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ar/guaranteed_possible_impossible_arht/results.json
diff --git a/evaluation_xnlimtht/xnli/ar/guaranteed_possible_impossible_armt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ar/guaranteed_possible_impossible_armt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/ar/guaranteed_possible_impossible_armt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ar/guaranteed_possible_impossible_armt/results.json
diff --git a/evaluation_xnlimtht/xnli/ar/justified_in_saying_arht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ar/justified_in_saying_arht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/ar/justified_in_saying_arht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ar/justified_in_saying_arht/results.json
diff --git a/evaluation_xnlimtht/xnli/ar/justified_in_saying_armt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ar/justified_in_saying_armt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/ar/justified_in_saying_armt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ar/justified_in_saying_armt/results.json
diff --git a/evaluation_xnlimtht/xnli/es/GPT-3_style_esht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/es/GPT-3_style_esht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/es/GPT-3_style_esht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/es/GPT-3_style_esht/results.json
diff --git a/evaluation_xnlimtht/xnli/es/GPT-3_style_esmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/es/GPT-3_style_esmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/es/GPT-3_style_esmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/es/GPT-3_style_esmt/results.json
diff --git a/evaluation_xnlimtht/xnli/es/MNLI_crowdsource_esht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/es/MNLI_crowdsource_esht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/es/MNLI_crowdsource_esht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/es/MNLI_crowdsource_esht/results.json
diff --git a/evaluation_xnlimtht/xnli/es/MNLI_crowdsource_esmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/es/MNLI_crowdsource_esmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/es/MNLI_crowdsource_esmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/es/MNLI_crowdsource_esmt/results.json
diff --git a/evaluation_xnlimtht/xnli/es/can_we_infer_esht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/es/can_we_infer_esht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/es/can_we_infer_esht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/es/can_we_infer_esht/results.json
diff --git a/evaluation_xnlimtht/xnli/es/can_we_infer_esmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/es/can_we_infer_esmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/es/can_we_infer_esmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/es/can_we_infer_esmt/results.json
diff --git a/evaluation_xnlimtht/xnli/es/guaranteed_possible_impossible_esht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/es/guaranteed_possible_impossible_esht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/es/guaranteed_possible_impossible_esht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/es/guaranteed_possible_impossible_esht/results.json
diff --git a/evaluation_xnlimtht/xnli/es/guaranteed_possible_impossible_esmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/es/guaranteed_possible_impossible_esmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/es/guaranteed_possible_impossible_esmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/es/guaranteed_possible_impossible_esmt/results.json
diff --git a/evaluation_xnlimtht/xnli/es/justified_in_saying_esht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/es/justified_in_saying_esht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/es/justified_in_saying_esht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/es/justified_in_saying_esht/results.json
diff --git a/evaluation_xnlimtht/xnli/es/justified_in_saying_esmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/es/justified_in_saying_esmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/es/justified_in_saying_esmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/es/justified_in_saying_esmt/results.json
diff --git a/evaluation_xnlimtht/xnli/fr/GPT-3_style_frht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/fr/GPT-3_style_frht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/fr/GPT-3_style_frht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/fr/GPT-3_style_frht/results.json
diff --git a/evaluation_xnlimtht/xnli/fr/GPT-3_style_frmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/fr/GPT-3_style_frmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/fr/GPT-3_style_frmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/fr/GPT-3_style_frmt/results.json
diff --git a/evaluation_xnlimtht/xnli/fr/MNLI_crowdsource_frht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/fr/MNLI_crowdsource_frht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/fr/MNLI_crowdsource_frht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/fr/MNLI_crowdsource_frht/results.json
diff --git a/evaluation_xnlimtht/xnli/fr/MNLI_crowdsource_frmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/fr/MNLI_crowdsource_frmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/fr/MNLI_crowdsource_frmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/fr/MNLI_crowdsource_frmt/results.json
diff --git a/evaluation_xnlimtht/xnli/fr/can_we_infer_frht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/fr/can_we_infer_frht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/fr/can_we_infer_frht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/fr/can_we_infer_frht/results.json
diff --git a/evaluation_xnlimtht/xnli/fr/can_we_infer_frmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/fr/can_we_infer_frmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/fr/can_we_infer_frmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/fr/can_we_infer_frmt/results.json
diff --git a/evaluation_xnlimtht/xnli/fr/guaranteed_possible_impossible_frht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/fr/guaranteed_possible_impossible_frht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/fr/guaranteed_possible_impossible_frht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/fr/guaranteed_possible_impossible_frht/results.json
diff --git a/evaluation_xnlimtht/xnli/fr/guaranteed_possible_impossible_frmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/fr/guaranteed_possible_impossible_frmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/fr/guaranteed_possible_impossible_frmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/fr/guaranteed_possible_impossible_frmt/results.json
diff --git a/evaluation_xnlimtht/xnli/fr/justified_in_saying_frht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/fr/justified_in_saying_frht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/fr/justified_in_saying_frht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/fr/justified_in_saying_frht/results.json
diff --git a/evaluation_xnlimtht/xnli/fr/justified_in_saying_frmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/fr/justified_in_saying_frmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/fr/justified_in_saying_frmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/fr/justified_in_saying_frmt/results.json
diff --git a/evaluation_xnlimtht/xnli/hi/GPT-3_style_hiht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/hi/GPT-3_style_hiht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/hi/GPT-3_style_hiht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/hi/GPT-3_style_hiht/results.json
diff --git a/evaluation_xnlimtht/xnli/hi/GPT-3_style_himt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/hi/GPT-3_style_himt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/hi/GPT-3_style_himt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/hi/GPT-3_style_himt/results.json
diff --git a/evaluation_xnlimtht/xnli/hi/MNLI_crowdsource_hiht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/hi/MNLI_crowdsource_hiht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/hi/MNLI_crowdsource_hiht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/hi/MNLI_crowdsource_hiht/results.json
diff --git a/evaluation_xnlimtht/xnli/hi/MNLI_crowdsource_himt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/hi/MNLI_crowdsource_himt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/hi/MNLI_crowdsource_himt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/hi/MNLI_crowdsource_himt/results.json
diff --git a/evaluation_xnlimtht/xnli/hi/can_we_infer_hiht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/hi/can_we_infer_hiht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/hi/can_we_infer_hiht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/hi/can_we_infer_hiht/results.json
diff --git a/evaluation_xnlimtht/xnli/hi/can_we_infer_himt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/hi/can_we_infer_himt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/hi/can_we_infer_himt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/hi/can_we_infer_himt/results.json
diff --git a/evaluation_xnlimtht/xnli/hi/guaranteed_possible_impossible_hiht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/hi/guaranteed_possible_impossible_hiht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/hi/guaranteed_possible_impossible_hiht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/hi/guaranteed_possible_impossible_hiht/results.json
diff --git a/evaluation_xnlimtht/xnli/hi/guaranteed_possible_impossible_himt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/hi/guaranteed_possible_impossible_himt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/hi/guaranteed_possible_impossible_himt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/hi/guaranteed_possible_impossible_himt/results.json
diff --git a/evaluation_xnlimtht/xnli/hi/justified_in_saying_hiht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/hi/justified_in_saying_hiht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/hi/justified_in_saying_hiht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/hi/justified_in_saying_hiht/results.json
diff --git a/evaluation_xnlimtht/xnli/hi/justified_in_saying_himt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/hi/justified_in_saying_himt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/hi/justified_in_saying_himt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/hi/justified_in_saying_himt/results.json
diff --git a/evaluation_xnlimtht/xnli/sw/GPT-3_style_swht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/sw/GPT-3_style_swht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/sw/GPT-3_style_swht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/sw/GPT-3_style_swht/results.json
diff --git a/evaluation_xnlimtht/xnli/sw/GPT-3_style_swmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/sw/GPT-3_style_swmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/sw/GPT-3_style_swmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/sw/GPT-3_style_swmt/results.json
diff --git a/evaluation_xnlimtht/xnli/sw/MNLI_crowdsource_swht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/sw/MNLI_crowdsource_swht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/sw/MNLI_crowdsource_swht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/sw/MNLI_crowdsource_swht/results.json
diff --git a/evaluation_xnlimtht/xnli/sw/MNLI_crowdsource_swmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/sw/MNLI_crowdsource_swmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/sw/MNLI_crowdsource_swmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/sw/MNLI_crowdsource_swmt/results.json
diff --git a/evaluation_xnlimtht/xnli/sw/can_we_infer_swht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/sw/can_we_infer_swht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/sw/can_we_infer_swht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/sw/can_we_infer_swht/results.json
diff --git a/evaluation_xnlimtht/xnli/sw/can_we_infer_swmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/sw/can_we_infer_swmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/sw/can_we_infer_swmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/sw/can_we_infer_swmt/results.json
diff --git a/evaluation_xnlimtht/xnli/sw/guaranteed_possible_impossible_swht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/sw/guaranteed_possible_impossible_swht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/sw/guaranteed_possible_impossible_swht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/sw/guaranteed_possible_impossible_swht/results.json
diff --git a/evaluation_xnlimtht/xnli/sw/guaranteed_possible_impossible_swmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/sw/guaranteed_possible_impossible_swmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/sw/guaranteed_possible_impossible_swmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/sw/guaranteed_possible_impossible_swmt/results.json
diff --git a/evaluation_xnlimtht/xnli/sw/justified_in_saying_swht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/sw/justified_in_saying_swht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/sw/justified_in_saying_swht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/sw/justified_in_saying_swht/results.json
diff --git a/evaluation_xnlimtht/xnli/sw/justified_in_saying_swmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/sw/justified_in_saying_swmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/sw/justified_in_saying_swmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/sw/justified_in_saying_swmt/results.json
diff --git a/evaluation_xnlimtht/xnli/ur/GPT-3_style_urht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ur/GPT-3_style_urht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/ur/GPT-3_style_urht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ur/GPT-3_style_urht/results.json
diff --git a/evaluation_xnlimtht/xnli/ur/GPT-3_style_urmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ur/GPT-3_style_urmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/ur/GPT-3_style_urmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ur/GPT-3_style_urmt/results.json
diff --git a/evaluation_xnlimtht/xnli/ur/MNLI_crowdsource_urht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ur/MNLI_crowdsource_urht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/ur/MNLI_crowdsource_urht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ur/MNLI_crowdsource_urht/results.json
diff --git a/evaluation_xnlimtht/xnli/ur/MNLI_crowdsource_urmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ur/MNLI_crowdsource_urmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/ur/MNLI_crowdsource_urmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ur/MNLI_crowdsource_urmt/results.json
diff --git a/evaluation_xnlimtht/xnli/ur/can_we_infer_urht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ur/can_we_infer_urht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/ur/can_we_infer_urht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ur/can_we_infer_urht/results.json
diff --git a/evaluation_xnlimtht/xnli/ur/can_we_infer_urmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ur/can_we_infer_urmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/ur/can_we_infer_urmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ur/can_we_infer_urmt/results.json
diff --git a/evaluation_xnlimtht/xnli/ur/guaranteed_possible_impossible_urht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ur/guaranteed_possible_impossible_urht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/ur/guaranteed_possible_impossible_urht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ur/guaranteed_possible_impossible_urht/results.json
diff --git a/evaluation_xnlimtht/xnli/ur/guaranteed_possible_impossible_urmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ur/guaranteed_possible_impossible_urmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/ur/guaranteed_possible_impossible_urmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ur/guaranteed_possible_impossible_urmt/results.json
diff --git a/evaluation_xnlimtht/xnli/ur/justified_in_saying_urht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ur/justified_in_saying_urht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/ur/justified_in_saying_urht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ur/justified_in_saying_urht/results.json
diff --git a/evaluation_xnlimtht/xnli/ur/justified_in_saying_urmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ur/justified_in_saying_urmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/ur/justified_in_saying_urmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/ur/justified_in_saying_urmt/results.json
diff --git a/evaluation_xnlimtht/xnli/vi/GPT-3_style_viht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/vi/GPT-3_style_viht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/vi/GPT-3_style_viht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/vi/GPT-3_style_viht/results.json
diff --git a/evaluation_xnlimtht/xnli/vi/GPT-3_style_vimt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/vi/GPT-3_style_vimt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/vi/GPT-3_style_vimt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/vi/GPT-3_style_vimt/results.json
diff --git a/evaluation_xnlimtht/xnli/vi/MNLI_crowdsource_viht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/vi/MNLI_crowdsource_viht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/vi/MNLI_crowdsource_viht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/vi/MNLI_crowdsource_viht/results.json
diff --git a/evaluation_xnlimtht/xnli/vi/MNLI_crowdsource_vimt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/vi/MNLI_crowdsource_vimt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/vi/MNLI_crowdsource_vimt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/vi/MNLI_crowdsource_vimt/results.json
diff --git a/evaluation_xnlimtht/xnli/vi/can_we_infer_viht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/vi/can_we_infer_viht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/vi/can_we_infer_viht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/vi/can_we_infer_viht/results.json
diff --git a/evaluation_xnlimtht/xnli/vi/can_we_infer_vimt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/vi/can_we_infer_vimt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/vi/can_we_infer_vimt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/vi/can_we_infer_vimt/results.json
diff --git a/evaluation_xnlimtht/xnli/vi/guaranteed_possible_impossible_viht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/vi/guaranteed_possible_impossible_viht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/vi/guaranteed_possible_impossible_viht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/vi/guaranteed_possible_impossible_viht/results.json
diff --git a/evaluation_xnlimtht/xnli/vi/guaranteed_possible_impossible_vimt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/vi/guaranteed_possible_impossible_vimt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/vi/guaranteed_possible_impossible_vimt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/vi/guaranteed_possible_impossible_vimt/results.json
diff --git a/evaluation_xnlimtht/xnli/vi/justified_in_saying_viht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/vi/justified_in_saying_viht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/vi/justified_in_saying_viht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/vi/justified_in_saying_viht/results.json
diff --git a/evaluation_xnlimtht/xnli/vi/justified_in_saying_vimt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/vi/justified_in_saying_vimt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/vi/justified_in_saying_vimt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/vi/justified_in_saying_vimt/results.json
diff --git a/evaluation_xnlimtht/xnli/zh/GPT-3_style_zhht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/zh/GPT-3_style_zhht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/zh/GPT-3_style_zhht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/zh/GPT-3_style_zhht/results.json
diff --git a/evaluation_xnlimtht/xnli/zh/GPT-3_style_zhmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/zh/GPT-3_style_zhmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/zh/GPT-3_style_zhmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/zh/GPT-3_style_zhmt/results.json
diff --git a/evaluation_xnlimtht/xnli/zh/MNLI_crowdsource_zhht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/zh/MNLI_crowdsource_zhht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/zh/MNLI_crowdsource_zhht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/zh/MNLI_crowdsource_zhht/results.json
diff --git a/evaluation_xnlimtht/xnli/zh/MNLI_crowdsource_zhmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/zh/MNLI_crowdsource_zhmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/zh/MNLI_crowdsource_zhmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/zh/MNLI_crowdsource_zhmt/results.json
diff --git a/evaluation_xnlimtht/xnli/zh/can_we_infer_zhht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/zh/can_we_infer_zhht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/zh/can_we_infer_zhht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/zh/can_we_infer_zhht/results.json
diff --git a/evaluation_xnlimtht/xnli/zh/can_we_infer_zhmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/zh/can_we_infer_zhmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/zh/can_we_infer_zhmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/zh/can_we_infer_zhmt/results.json
diff --git a/evaluation_xnlimtht/xnli/zh/guaranteed_possible_impossible_zhht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/zh/guaranteed_possible_impossible_zhht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/zh/guaranteed_possible_impossible_zhht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/zh/guaranteed_possible_impossible_zhht/results.json
diff --git a/evaluation_xnlimtht/xnli/zh/guaranteed_possible_impossible_zhmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/zh/guaranteed_possible_impossible_zhmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/zh/guaranteed_possible_impossible_zhmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/zh/guaranteed_possible_impossible_zhmt/results.json
diff --git a/evaluation_xnlimtht/xnli/zh/justified_in_saying_zhht/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/zh/justified_in_saying_zhht/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/zh/justified_in_saying_zhht/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/zh/justified_in_saying_zhht/results.json
diff --git a/evaluation_xnlimtht/xnli/zh/justified_in_saying_zhmt/results.json b/evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/zh/justified_in_saying_zhmt/results.json
similarity index 100%
rename from evaluation_xnlimtht/xnli/zh/justified_in_saying_zhmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xnlimtht/xnli/zh/justified_in_saying_zhmt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/ar/Answer_Given_options_armt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/ar/Answer_Given_options_armt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/ar/Answer_Given_options_armt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/ar/Answer_Given_options_armt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/ar/Choose_Story_Ending_armt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/ar/Choose_Story_Ending_armt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/ar/Choose_Story_Ending_armt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/ar/Choose_Story_Ending_armt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/ar/Generate_Ending_armt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/ar/Generate_Ending_armt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/ar/Generate_Ending_armt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/ar/Generate_Ending_armt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/ar/Novel_Correct_Ending_armt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/ar/Novel_Correct_Ending_armt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/ar/Novel_Correct_Ending_armt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/ar/Novel_Correct_Ending_armt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/ar/Story_Continuation_and_Options_armt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/ar/Story_Continuation_and_Options_armt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/ar/Story_Continuation_and_Options_armt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/ar/Story_Continuation_and_Options_armt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/es/Answer_Given_options_esmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/es/Answer_Given_options_esmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/es/Answer_Given_options_esmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/es/Answer_Given_options_esmt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/es/Choose_Story_Ending_esmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/es/Choose_Story_Ending_esmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/es/Choose_Story_Ending_esmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/es/Choose_Story_Ending_esmt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/es/Generate_Ending_esmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/es/Generate_Ending_esmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/es/Generate_Ending_esmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/es/Generate_Ending_esmt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/es/Novel_Correct_Ending_esmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/es/Novel_Correct_Ending_esmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/es/Novel_Correct_Ending_esmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/es/Novel_Correct_Ending_esmt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/es/Story_Continuation_and_Options_esmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/es/Story_Continuation_and_Options_esmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/es/Story_Continuation_and_Options_esmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/es/Story_Continuation_and_Options_esmt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/eu/Answer_Given_options_eumt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/eu/Answer_Given_options_eumt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/eu/Answer_Given_options_eumt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/eu/Answer_Given_options_eumt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/eu/Choose_Story_Ending_eumt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/eu/Choose_Story_Ending_eumt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/eu/Choose_Story_Ending_eumt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/eu/Choose_Story_Ending_eumt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/eu/Generate_Ending_eumt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/eu/Generate_Ending_eumt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/eu/Generate_Ending_eumt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/eu/Generate_Ending_eumt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/eu/Novel_Correct_Ending_eumt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/eu/Novel_Correct_Ending_eumt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/eu/Novel_Correct_Ending_eumt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/eu/Novel_Correct_Ending_eumt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/eu/Story_Continuation_and_Options_eumt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/eu/Story_Continuation_and_Options_eumt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/eu/Story_Continuation_and_Options_eumt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/eu/Story_Continuation_and_Options_eumt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/hi/Answer_Given_options_himt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/hi/Answer_Given_options_himt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/hi/Answer_Given_options_himt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/hi/Answer_Given_options_himt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/hi/Choose_Story_Ending_himt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/hi/Choose_Story_Ending_himt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/hi/Choose_Story_Ending_himt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/hi/Choose_Story_Ending_himt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/hi/Generate_Ending_himt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/hi/Generate_Ending_himt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/hi/Generate_Ending_himt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/hi/Generate_Ending_himt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/hi/Novel_Correct_Ending_himt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/hi/Novel_Correct_Ending_himt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/hi/Novel_Correct_Ending_himt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/hi/Novel_Correct_Ending_himt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/hi/Story_Continuation_and_Options_himt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/hi/Story_Continuation_and_Options_himt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/hi/Story_Continuation_and_Options_himt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/hi/Story_Continuation_and_Options_himt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/id/Answer_Given_options_idmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/id/Answer_Given_options_idmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/id/Answer_Given_options_idmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/id/Answer_Given_options_idmt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/id/Choose_Story_Ending_idmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/id/Choose_Story_Ending_idmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/id/Choose_Story_Ending_idmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/id/Choose_Story_Ending_idmt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/id/Generate_Ending_idmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/id/Generate_Ending_idmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/id/Generate_Ending_idmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/id/Generate_Ending_idmt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/id/Novel_Correct_Ending_idmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/id/Novel_Correct_Ending_idmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/id/Novel_Correct_Ending_idmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/id/Novel_Correct_Ending_idmt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/id/Story_Continuation_and_Options_idmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/id/Story_Continuation_and_Options_idmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/id/Story_Continuation_and_Options_idmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/id/Story_Continuation_and_Options_idmt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/zh/Answer_Given_options_zhmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/zh/Answer_Given_options_zhmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/zh/Answer_Given_options_zhmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/zh/Answer_Given_options_zhmt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/zh/Choose_Story_Ending_zhmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/zh/Choose_Story_Ending_zhmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/zh/Choose_Story_Ending_zhmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/zh/Choose_Story_Ending_zhmt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/zh/Generate_Ending_zhmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/zh/Generate_Ending_zhmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/zh/Generate_Ending_zhmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/zh/Generate_Ending_zhmt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/zh/Novel_Correct_Ending_zhmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/zh/Novel_Correct_Ending_zhmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/zh/Novel_Correct_Ending_zhmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/zh/Novel_Correct_Ending_zhmt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/zh/Story_Continuation_and_Options_zhmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/zh/Story_Continuation_and_Options_zhmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/zh/Story_Continuation_and_Options_zhmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xstory_cloze/zh/Story_Continuation_and_Options_zhmt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xwinograd/fr/Replace_frmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xwinograd/fr/Replace_frmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xwinograd/fr/Replace_frmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xwinograd/fr/Replace_frmt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xwinograd/fr/True_or_False_frmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xwinograd/fr/True_or_False_frmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xwinograd/fr/True_or_False_frmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xwinograd/fr/True_or_False_frmt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xwinograd/fr/does_underscore_refer_to_frmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xwinograd/fr/does_underscore_refer_to_frmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xwinograd/fr/does_underscore_refer_to_frmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xwinograd/fr/does_underscore_refer_to_frmt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xwinograd/fr/stand_for_frmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xwinograd/fr/stand_for_frmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xwinograd/fr/stand_for_frmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xwinograd/fr/stand_for_frmt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xwinograd/fr/underscore_refer_to_frmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xwinograd/fr/underscore_refer_to_frmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xwinograd/fr/underscore_refer_to_frmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xwinograd/fr/underscore_refer_to_frmt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xwinograd/pt/Replace_ptmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xwinograd/pt/Replace_ptmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xwinograd/pt/Replace_ptmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xwinograd/pt/Replace_ptmt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xwinograd/pt/True_or_False_ptmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xwinograd/pt/True_or_False_ptmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xwinograd/pt/True_or_False_ptmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xwinograd/pt/True_or_False_ptmt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xwinograd/pt/does_underscore_refer_to_ptmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xwinograd/pt/does_underscore_refer_to_ptmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xwinograd/pt/does_underscore_refer_to_ptmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xwinograd/pt/does_underscore_refer_to_ptmt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xwinograd/pt/stand_for_ptmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xwinograd/pt/stand_for_ptmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xwinograd/pt/stand_for_ptmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xwinograd/pt/stand_for_ptmt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xwinograd/pt/underscore_refer_to_ptmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xwinograd/pt/underscore_refer_to_ptmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xwinograd/pt/underscore_refer_to_ptmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xwinograd/pt/underscore_refer_to_ptmt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xwinograd/zh/Replace_zhmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xwinograd/zh/Replace_zhmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xwinograd/zh/Replace_zhmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xwinograd/zh/Replace_zhmt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xwinograd/zh/True_or_False_zhmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xwinograd/zh/True_or_False_zhmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xwinograd/zh/True_or_False_zhmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xwinograd/zh/True_or_False_zhmt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xwinograd/zh/does_underscore_refer_to_zhmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xwinograd/zh/does_underscore_refer_to_zhmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xwinograd/zh/does_underscore_refer_to_zhmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xwinograd/zh/does_underscore_refer_to_zhmt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xwinograd/zh/stand_for_zhmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xwinograd/zh/stand_for_zhmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xwinograd/zh/stand_for_zhmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xwinograd/zh/stand_for_zhmt/results.json
diff --git a/evaluation_xwinostorycopamt/Muennighoff_xwinograd/zh/underscore_refer_to_zhmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xwinograd/zh/underscore_refer_to_zhmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/Muennighoff_xwinograd/zh/underscore_refer_to_zhmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/Muennighoff_xwinograd/zh/underscore_refer_to_zhmt/results.json
diff --git a/evaluation_xwinostorycopamt/xcopa/id/C1_or_C2?_premise_idmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/id/C1_or_C2?_premise_idmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/xcopa/id/C1_or_C2?_premise_idmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/id/C1_or_C2?_premise_idmt/results.json
diff --git a/evaluation_xwinostorycopamt/xcopa/id/best_option_idmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/id/best_option_idmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/xcopa/id/best_option_idmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/id/best_option_idmt/results.json
diff --git a/evaluation_xwinostorycopamt/xcopa/id/cause_effect_idmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/id/cause_effect_idmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/xcopa/id/cause_effect_idmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/id/cause_effect_idmt/results.json
diff --git a/evaluation_xwinostorycopamt/xcopa/id/i_am_hesitating_idmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/id/i_am_hesitating_idmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/xcopa/id/i_am_hesitating_idmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/id/i_am_hesitating_idmt/results.json
diff --git a/evaluation_xwinostorycopamt/xcopa/id/plausible_alternatives_idmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/id/plausible_alternatives_idmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/xcopa/id/plausible_alternatives_idmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/id/plausible_alternatives_idmt/results.json
diff --git a/evaluation_xwinostorycopamt/xcopa/sw/C1_or_C2?_premise_swmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/sw/C1_or_C2?_premise_swmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/xcopa/sw/C1_or_C2?_premise_swmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/sw/C1_or_C2?_premise_swmt/results.json
diff --git a/evaluation_xwinostorycopamt/xcopa/sw/best_option_swmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/sw/best_option_swmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/xcopa/sw/best_option_swmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/sw/best_option_swmt/results.json
diff --git a/evaluation_xwinostorycopamt/xcopa/sw/cause_effect_swmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/sw/cause_effect_swmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/xcopa/sw/cause_effect_swmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/sw/cause_effect_swmt/results.json
diff --git a/evaluation_xwinostorycopamt/xcopa/sw/i_am_hesitating_swmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/sw/i_am_hesitating_swmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/xcopa/sw/i_am_hesitating_swmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/sw/i_am_hesitating_swmt/results.json
diff --git a/evaluation_xwinostorycopamt/xcopa/sw/plausible_alternatives_swmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/sw/plausible_alternatives_swmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/xcopa/sw/plausible_alternatives_swmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/sw/plausible_alternatives_swmt/results.json
diff --git a/evaluation_xwinostorycopamt/xcopa/ta/C1_or_C2?_premise_tamt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/ta/C1_or_C2?_premise_tamt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/xcopa/ta/C1_or_C2?_premise_tamt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/ta/C1_or_C2?_premise_tamt/results.json
diff --git a/evaluation_xwinostorycopamt/xcopa/ta/best_option_tamt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/ta/best_option_tamt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/xcopa/ta/best_option_tamt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/ta/best_option_tamt/results.json
diff --git a/evaluation_xwinostorycopamt/xcopa/ta/cause_effect_tamt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/ta/cause_effect_tamt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/xcopa/ta/cause_effect_tamt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/ta/cause_effect_tamt/results.json
diff --git a/evaluation_xwinostorycopamt/xcopa/ta/i_am_hesitating_tamt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/ta/i_am_hesitating_tamt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/xcopa/ta/i_am_hesitating_tamt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/ta/i_am_hesitating_tamt/results.json
diff --git a/evaluation_xwinostorycopamt/xcopa/ta/plausible_alternatives_tamt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/ta/plausible_alternatives_tamt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/xcopa/ta/plausible_alternatives_tamt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/ta/plausible_alternatives_tamt/results.json
diff --git a/evaluation_xwinostorycopamt/xcopa/vi/C1_or_C2?_premise_vimt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/vi/C1_or_C2?_premise_vimt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/xcopa/vi/C1_or_C2?_premise_vimt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/vi/C1_or_C2?_premise_vimt/results.json
diff --git a/evaluation_xwinostorycopamt/xcopa/vi/best_option_vimt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/vi/best_option_vimt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/xcopa/vi/best_option_vimt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/vi/best_option_vimt/results.json
diff --git a/evaluation_xwinostorycopamt/xcopa/vi/cause_effect_vimt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/vi/cause_effect_vimt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/xcopa/vi/cause_effect_vimt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/vi/cause_effect_vimt/results.json
diff --git a/evaluation_xwinostorycopamt/xcopa/vi/i_am_hesitating_vimt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/vi/i_am_hesitating_vimt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/xcopa/vi/i_am_hesitating_vimt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/vi/i_am_hesitating_vimt/results.json
diff --git a/evaluation_xwinostorycopamt/xcopa/vi/plausible_alternatives_vimt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/vi/plausible_alternatives_vimt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/xcopa/vi/plausible_alternatives_vimt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/vi/plausible_alternatives_vimt/results.json
diff --git a/evaluation_xwinostorycopamt/xcopa/zh/C1_or_C2?_premise_zhmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/zh/C1_or_C2?_premise_zhmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/xcopa/zh/C1_or_C2?_premise_zhmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/zh/C1_or_C2?_premise_zhmt/results.json
diff --git a/evaluation_xwinostorycopamt/xcopa/zh/best_option_zhmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/zh/best_option_zhmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/xcopa/zh/best_option_zhmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/zh/best_option_zhmt/results.json
diff --git a/evaluation_xwinostorycopamt/xcopa/zh/cause_effect_zhmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/zh/cause_effect_zhmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/xcopa/zh/cause_effect_zhmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/zh/cause_effect_zhmt/results.json
diff --git a/evaluation_xwinostorycopamt/xcopa/zh/i_am_hesitating_zhmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/zh/i_am_hesitating_zhmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/xcopa/zh/i_am_hesitating_zhmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/zh/i_am_hesitating_zhmt/results.json
diff --git a/evaluation_xwinostorycopamt/xcopa/zh/plausible_alternatives_zhmt/results.json b/evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/zh/plausible_alternatives_zhmt/results.json
similarity index 100%
rename from evaluation_xwinostorycopamt/xcopa/zh/plausible_alternatives_zhmt/results.json
rename to evaluation_bloomz-1b7/evaluation_xwinostorycopamt/xcopa/zh/plausible_alternatives_zhmt/results.json