diff --git a/evaluation_val/amazon_reviews_multi/en/prompt_body_title_to_star/results.json b/evaluation_bloomz-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/evaluation_val/aqua_rat/raw/select_the_best_option/results.json
diff --git a/evaluation_val/art/choose_hypothesis/results.json b/evaluation_bloomz-7b1/evaluation_val/art/choose_hypothesis/results.json
similarity index 100%
rename from evaluation_val/art/choose_hypothesis/results.json
rename to evaluation_bloomz-7b1/evaluation_val/art/choose_hypothesis/results.json
diff --git a/evaluation_val/art/choose_hypothesis_believable/results.json b/evaluation_bloomz-7b1/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-7b1/evaluation_val/art/choose_hypothesis_believable/results.json
diff --git a/evaluation_val/art/choose_hypothesis_desc/results.json b/evaluation_bloomz-7b1/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-7b1/evaluation_val/art/choose_hypothesis_desc/results.json
diff --git a/evaluation_val/art/choose_hypothesis_likely/results.json b/evaluation_bloomz-7b1/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-7b1/evaluation_val/art/choose_hypothesis_likely/results.json
diff --git a/evaluation_val/art/choose_hypothesis_options/results.json b/evaluation_bloomz-7b1/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-7b1/evaluation_val/art/choose_hypothesis_options/results.json
diff --git a/evaluation_val/banking77/direct_to_which_department/results.json b/evaluation_bloomz-7b1/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-7b1/evaluation_val/banking77/direct_to_which_department/results.json
diff --git a/evaluation_val/banking77/help_page_topic/results.json b/evaluation_bloomz-7b1/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-7b1/evaluation_val/banking77/help_page_topic/results.json
diff --git a/evaluation_val/banking77/rephrase_as_banking_term/results.json b/evaluation_bloomz-7b1/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-7b1/evaluation_val/banking77/rephrase_as_banking_term/results.json
diff --git a/evaluation_val/blbooksgenre/title_genre_classifiction/classify/results.json b/evaluation_bloomz-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/evaluation_val/commonsense_qa/most_suitable_answer/results.json
diff --git a/evaluation_val/commonsense_qa/question_answering/results.json b/evaluation_bloomz-7b1/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-7b1/evaluation_val/commonsense_qa/question_answering/results.json
diff --git a/evaluation_val/conv_ai_3/ambiguous/results.json b/evaluation_bloomz-7b1/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-7b1/evaluation_val/conv_ai_3/ambiguous/results.json
diff --git a/evaluation_val/conv_ai_3/clarification_needed/results.json b/evaluation_bloomz-7b1/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-7b1/evaluation_val/conv_ai_3/clarification_needed/results.json
diff --git a/evaluation_val/conv_ai_3/directly_answer/results.json b/evaluation_bloomz-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/evaluation_val/conv_ai_3/score_how_much/results.json
diff --git a/evaluation_val/craigslist_bargains/best_deal/results.json b/evaluation_bloomz-7b1/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-7b1/evaluation_val/craigslist_bargains/best_deal/results.json
diff --git a/evaluation_val/craigslist_bargains/good_deal_for_seller/results.json b/evaluation_bloomz-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/evaluation_val/emotion/reply_with_emoation_label/results.json
diff --git a/evaluation_val/evaluation_mlsum/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=layman_summ_es.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:39.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_mlsum/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=layman_summ_es.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:39.json
similarity index 100%
rename from evaluation_val/evaluation_mlsum/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=layman_summ_es.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:39.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_mlsum/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=layman_summ_es.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:39.json
diff --git a/evaluation_val/evaluation_mlsum/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=palm_prompt.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:39.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_mlsum/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=palm_prompt.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:39.json
similarity index 100%
rename from evaluation_val/evaluation_mlsum/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=palm_prompt.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:39.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_mlsum/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=palm_prompt.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:39.json
diff --git a/evaluation_val/evaluation_mlsum/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=summarise_this_in_es_few_sentences.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:37.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_mlsum/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=summarise_this_in_es_few_sentences.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:37.json
similarity index 100%
rename from evaluation_val/evaluation_mlsum/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=summarise_this_in_es_few_sentences.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:37.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_mlsum/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=summarise_this_in_es_few_sentences.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:37.json
diff --git a/evaluation_val/evaluation_mlsum/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=layman_summ_es.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:39.jsonl b/evaluation_bloomz-7b1/evaluation_val/evaluation_mlsum/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=layman_summ_es.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:39.jsonl
similarity index 100%
rename from evaluation_val/evaluation_mlsum/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=layman_summ_es.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:39.jsonl
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_mlsum/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=layman_summ_es.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:39.jsonl
diff --git a/evaluation_val/evaluation_mlsum/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=palm_prompt.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:39.jsonl b/evaluation_bloomz-7b1/evaluation_val/evaluation_mlsum/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=palm_prompt.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:39.jsonl
similarity index 100%
rename from evaluation_val/evaluation_mlsum/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=palm_prompt.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:39.jsonl
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_mlsum/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=palm_prompt.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:39.jsonl
diff --git a/evaluation_val/evaluation_mlsum/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=summarise_this_in_es_few_sentences.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:37.jsonl b/evaluation_bloomz-7b1/evaluation_val/evaluation_mlsum/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=summarise_this_in_es_few_sentences.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:37.jsonl
similarity index 100%
rename from evaluation_val/evaluation_mlsum/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=summarise_this_in_es_few_sentences.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:37.jsonl
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_mlsum/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=summarise_this_in_es_few_sentences.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:37.jsonl
diff --git a/evaluation_val/evaluation_mlsum/merged.csv b/evaluation_bloomz-7b1/evaluation_val/evaluation_mlsum/merged.csv
similarity index 100%
rename from evaluation_val/evaluation_mlsum/merged.csv
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_mlsum/merged.csv
diff --git a/evaluation_val/evaluation_mlsum/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=layman_summ_es.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:39.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_mlsum/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=layman_summ_es.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:39.json
similarity index 100%
rename from evaluation_val/evaluation_mlsum/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=layman_summ_es.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:39.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_mlsum/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=layman_summ_es.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:39.json
diff --git a/evaluation_val/evaluation_mlsum/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=palm_prompt.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:39.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_mlsum/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=palm_prompt.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:39.json
similarity index 100%
rename from evaluation_val/evaluation_mlsum/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=palm_prompt.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:39.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_mlsum/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=palm_prompt.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:39.json
diff --git a/evaluation_val/evaluation_mlsum/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=summarise_this_in_es_few_sentences.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:37.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_mlsum/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=summarise_this_in_es_few_sentences.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:37.json
similarity index 100%
rename from evaluation_val/evaluation_mlsum/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=summarise_this_in_es_few_sentences.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:37.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_mlsum/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=mlsum_es.templates=summarise_this_in_es_few_sentences.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:04:37.json
diff --git a/evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-en-hi-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-en-hi-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.json
similarity index 100%
rename from evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-en-hi-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-en-hi-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.json
diff --git a/evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.json
similarity index 100%
rename from evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.json
diff --git a/evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-hi-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-hi-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.json
similarity index 100%
rename from evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-hi-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-hi-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.json
diff --git a/evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.json
similarity index 100%
rename from evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.json
diff --git a/evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt-3-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:37:50.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt-3-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:37:50.json
similarity index 100%
rename from evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt-3-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:37:50.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt-3-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:37:50.json
diff --git a/evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt-3-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:37:50.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt-3-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:37:50.json
similarity index 100%
rename from evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt-3-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:37:50.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt-3-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:37:50.json
diff --git a/evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=version-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=version-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.json
similarity index 100%
rename from evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=version-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=version-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.json
diff --git a/evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=version-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=version-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.json
similarity index 100%
rename from evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=version-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=version-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.json
diff --git a/evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=xglm-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=xglm-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.json
similarity index 100%
rename from evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=xglm-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=xglm-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.json
diff --git a/evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=xglm-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=xglm-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.json
similarity index 100%
rename from evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=xglm-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=xglm-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.json
diff --git a/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-en-hi-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.jsonl b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-en-hi-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.jsonl
similarity index 100%
rename from evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-en-hi-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.jsonl
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-en-hi-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.jsonl
diff --git a/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.jsonl b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.jsonl
similarity index 100%
rename from evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.jsonl
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.jsonl
diff --git a/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-hi-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.jsonl b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-hi-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.jsonl
similarity index 100%
rename from evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-hi-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.jsonl
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-hi-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.jsonl
diff --git a/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.jsonl b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.jsonl
similarity index 100%
rename from evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.jsonl
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.jsonl
diff --git a/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt-3-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:37:50.jsonl b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt-3-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:37:50.jsonl
similarity index 100%
rename from evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt-3-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:37:50.jsonl
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt-3-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:37:50.jsonl
diff --git a/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt-3-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:37:50.jsonl b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt-3-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:37:50.jsonl
similarity index 100%
rename from evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt-3-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:37:50.jsonl
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt-3-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:37:50.jsonl
diff --git a/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt3-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:25:09.jsonl b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt3-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:25:09.jsonl
similarity index 100%
rename from evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt3-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:25:09.jsonl
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt3-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:25:09.jsonl
diff --git a/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt3-en-hi.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.jsonl b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt3-en-hi.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.jsonl
similarity index 100%
rename from evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt3-en-hi.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.jsonl
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt3-en-hi.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.jsonl
diff --git a/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt3-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:25:09.jsonl b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt3-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:25:09.jsonl
similarity index 100%
rename from evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt3-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:25:09.jsonl
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt3-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:25:09.jsonl
diff --git a/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt3-hi-en.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.jsonl b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt3-hi-en.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.jsonl
similarity index 100%
rename from evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt3-hi-en.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.jsonl
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt3-hi-en.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.jsonl
diff --git a/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=version-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.jsonl b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=version-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.jsonl
similarity index 100%
rename from evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=version-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.jsonl
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=version-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.jsonl
diff --git a/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=version-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.jsonl b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=version-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.jsonl
similarity index 100%
rename from evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=version-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.jsonl
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=version-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.jsonl
diff --git a/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=xglm-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.jsonl b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=xglm-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.jsonl
similarity index 100%
rename from evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=xglm-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.jsonl
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=xglm-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.jsonl
diff --git a/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=xglm-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.jsonl b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=xglm-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.jsonl
similarity index 100%
rename from evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=xglm-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.jsonl
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=xglm-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.jsonl
diff --git a/evaluation_val/evaluation_wmt14_hi_en/merged.csv b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/merged.csv
similarity index 100%
rename from evaluation_val/evaluation_wmt14_hi_en/merged.csv
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/merged.csv
diff --git a/evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-en-hi-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-en-hi-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.json
similarity index 100%
rename from evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-en-hi-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-en-hi-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.json
diff --git a/evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.json
similarity index 100%
rename from evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.json
diff --git a/evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-hi-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-hi-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.json
similarity index 100%
rename from evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-hi-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-hi-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.json
diff --git a/evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.json
similarity index 100%
rename from evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=a_good_translation-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.json
diff --git a/evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt-3-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:37:50.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt-3-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:37:50.json
similarity index 100%
rename from evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt-3-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:37:50.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt-3-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:37:50.json
diff --git a/evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt-3-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:37:50.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt-3-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:37:50.json
similarity index 100%
rename from evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt-3-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:37:50.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=gpt-3-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T18:37:50.json
diff --git a/evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=version-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=version-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.json
similarity index 100%
rename from evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=version-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=version-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.json
diff --git a/evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=version-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=version-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.json
similarity index 100%
rename from evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=version-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=version-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.json
diff --git a/evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=xglm-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=xglm-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.json
similarity index 100%
rename from evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=xglm-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=xglm-en-hi-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:23.json
diff --git a/evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=xglm-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=xglm-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.json
similarity index 100%
rename from evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=xglm-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt14_hi_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_hi_en.templates=xglm-hi-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T17:51:24.json
diff --git a/evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-en-fr-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-en-fr-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
similarity index 100%
rename from evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-en-fr-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-en-fr-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
diff --git a/evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
similarity index 100%
rename from evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
diff --git a/evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-fr-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-fr-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.json
similarity index 100%
rename from evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-fr-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-fr-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.json
diff --git a/evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.json
similarity index 100%
rename from evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.json
diff --git a/evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=gpt3-en-fr.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=gpt3-en-fr.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
similarity index 100%
rename from evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=gpt3-en-fr.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=gpt3-en-fr.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
diff --git a/evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=gpt3-fr-en.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=gpt3-fr-en.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.json
similarity index 100%
rename from evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=gpt3-fr-en.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=gpt3-fr-en.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.json
diff --git a/evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=version-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=version-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
similarity index 100%
rename from evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=version-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=version-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
diff --git a/evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=version-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=version-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
similarity index 100%
rename from evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=version-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=version-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
diff --git a/evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=xglm-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=xglm-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
similarity index 100%
rename from evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=xglm-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=xglm-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
diff --git a/evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=xglm-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=xglm-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.json
similarity index 100%
rename from evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=xglm-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/agg.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=xglm-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.json
diff --git a/evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-en-fr-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.jsonl b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-en-fr-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.jsonl
similarity index 100%
rename from evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-en-fr-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.jsonl
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-en-fr-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.jsonl
diff --git a/evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.jsonl b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.jsonl
similarity index 100%
rename from evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.jsonl
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.jsonl
diff --git a/evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-fr-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.jsonl b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-fr-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.jsonl
similarity index 100%
rename from evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-fr-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.jsonl
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-fr-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.jsonl
diff --git a/evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.jsonl b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.jsonl
similarity index 100%
rename from evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.jsonl
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.jsonl
diff --git a/evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=gpt3-en-fr.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.jsonl b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=gpt3-en-fr.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.jsonl
similarity index 100%
rename from evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=gpt3-en-fr.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.jsonl
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=gpt3-en-fr.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.jsonl
diff --git a/evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=gpt3-fr-en.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.jsonl b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=gpt3-fr-en.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.jsonl
similarity index 100%
rename from evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=gpt3-fr-en.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.jsonl
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=gpt3-fr-en.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.jsonl
diff --git a/evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=version-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.jsonl b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=version-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.jsonl
similarity index 100%
rename from evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=version-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.jsonl
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=version-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.jsonl
diff --git a/evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=version-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.jsonl b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=version-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.jsonl
similarity index 100%
rename from evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=version-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.jsonl
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=version-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.jsonl
diff --git a/evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=xglm-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.jsonl b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=xglm-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.jsonl
similarity index 100%
rename from evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=xglm-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.jsonl
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=xglm-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.jsonl
diff --git a/evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=xglm-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.jsonl b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=xglm-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.jsonl
similarity index 100%
rename from evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=xglm-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.jsonl
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/examples.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=xglm-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.jsonl
diff --git a/evaluation_val/evaluation_wmt_fr_en/merged.csv b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/merged.csv
similarity index 100%
rename from evaluation_val/evaluation_wmt_fr_en/merged.csv
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/merged.csv
diff --git a/evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-en-fr-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-en-fr-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
similarity index 100%
rename from evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-en-fr-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-en-fr-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
diff --git a/evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
similarity index 100%
rename from evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
diff --git a/evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-fr-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-fr-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.json
similarity index 100%
rename from evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-fr-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-fr-en-source+target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.json
diff --git a/evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.json
similarity index 100%
rename from evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=a_good_translation-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.json
diff --git a/evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=gpt3-en-fr.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=gpt3-en-fr.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
similarity index 100%
rename from evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=gpt3-en-fr.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=gpt3-en-fr.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
diff --git a/evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=gpt3-fr-en.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=gpt3-fr-en.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.json
similarity index 100%
rename from evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=gpt3-fr-en.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=gpt3-fr-en.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.json
diff --git a/evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=version-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=version-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
similarity index 100%
rename from evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=version-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=version-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
diff --git a/evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=version-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=version-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
similarity index 100%
rename from evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=version-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=version-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
diff --git a/evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=xglm-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=xglm-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
similarity index 100%
rename from evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=xglm-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=xglm-en-fr-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:42.json
diff --git a/evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=xglm-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.json b/evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=xglm-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.json
similarity index 100%
rename from evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=xglm-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.json
rename to evaluation_bloomz-7b1/evaluation_val/evaluation_wmt_fr_en/slim.limited=3000.model=tr13f-6b3-ml-t0-lmtoks341b-t0toks4b2-xp3capmixnewcodelonglossseq.task=wmt14_fr_en.templates=xglm-fr-en-target.fewshot=0.batchsize=16.seed=1234.timestamp=2022-09-06T16:38:47.json
diff --git a/evaluation_val/financial_phrasebank/sentences_allagree/bullish_neutral_bearish/results.json b/evaluation_bloomz-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/evaluation_val/glue/cola/Previous_sentence_acceptable/results.json
diff --git a/evaluation_val/glue/cola/editing/results.json b/evaluation_bloomz-7b1/evaluation_val/glue/cola/editing/results.json
similarity index 100%
rename from evaluation_val/glue/cola/editing/results.json
rename to evaluation_bloomz-7b1/evaluation_val/glue/cola/editing/results.json
diff --git a/evaluation_val/glue/cola/is_this_correct/results.json b/evaluation_bloomz-7b1/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-7b1/evaluation_val/glue/cola/is_this_correct/results.json
diff --git a/evaluation_val/glue/sst2/following_positive_negative/results.json b/evaluation_bloomz-7b1/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-7b1/evaluation_val/glue/sst2/following_positive_negative/results.json
diff --git a/evaluation_val/glue/sst2/happy_or_mad/results.json b/evaluation_bloomz-7b1/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-7b1/evaluation_val/glue/sst2/happy_or_mad/results.json
diff --git a/evaluation_val/glue/sst2/positive_negative_after/results.json b/evaluation_bloomz-7b1/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-7b1/evaluation_val/glue/sst2/positive_negative_after/results.json
diff --git a/evaluation_val/glue/sst2/review/results.json b/evaluation_bloomz-7b1/evaluation_val/glue/sst2/review/results.json
similarity index 100%
rename from evaluation_val/glue/sst2/review/results.json
rename to evaluation_bloomz-7b1/evaluation_val/glue/sst2/review/results.json
diff --git a/evaluation_val/glue/sst2/said/results.json b/evaluation_bloomz-7b1/evaluation_val/glue/sst2/said/results.json
similarity index 100%
rename from evaluation_val/glue/sst2/said/results.json
rename to evaluation_bloomz-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/evaluation_val/liar/Given_statement_guess_category/results.json
diff --git a/evaluation_val/lince/sa_spaeng/express_sentiment/results.json b/evaluation_bloomz-7b1/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-7b1/evaluation_val/lince/sa_spaeng/express_sentiment/results.json
diff --git a/evaluation_val/lince/sa_spaeng/negation_template/results.json b/evaluation_bloomz-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/evaluation_val/math_qa/first_choice_then_problem/results.json
diff --git a/evaluation_val/math_qa/gre_problem/results.json b/evaluation_bloomz-7b1/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-7b1/evaluation_val/math_qa/gre_problem/results.json
diff --git a/evaluation_val/math_qa/pick_the_correct/results.json b/evaluation_bloomz-7b1/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-7b1/evaluation_val/math_qa/pick_the_correct/results.json
diff --git a/evaluation_val/math_qa/problem_set_type/results.json b/evaluation_bloomz-7b1/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-7b1/evaluation_val/math_qa/problem_set_type/results.json
diff --git a/evaluation_val/merged.csv b/evaluation_bloomz-7b1/evaluation_val/merged.csv
similarity index 100%
rename from evaluation_val/merged.csv
rename to evaluation_bloomz-7b1/evaluation_val/merged.csv
diff --git a/evaluation_val/merged.json b/evaluation_bloomz-7b1/evaluation_val/merged.json
similarity index 100%
rename from evaluation_val/merged.json
rename to evaluation_bloomz-7b1/evaluation_val/merged.json
diff --git a/evaluation_val/movie_rationales/Evidences_+_review/results.json b/evaluation_bloomz-7b1/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-7b1/evaluation_val/movie_rationales/Evidences_+_review/results.json
diff --git a/evaluation_val/movie_rationales/Evidences_sentiment_classification/results.json b/evaluation_bloomz-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/evaluation_val/mwsc/in-the-sentence-question-first/results.json
diff --git a/evaluation_val/mwsc/in-the-sentence/results.json b/evaluation_bloomz-7b1/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-7b1/evaluation_val/mwsc/in-the-sentence/results.json
diff --git a/evaluation_val/mwsc/is-correct/results.json b/evaluation_bloomz-7b1/evaluation_val/mwsc/is-correct/results.json
similarity index 100%
rename from evaluation_val/mwsc/is-correct/results.json
rename to evaluation_bloomz-7b1/evaluation_val/mwsc/is-correct/results.json
diff --git a/evaluation_val/mwsc/options-or/results.json b/evaluation_bloomz-7b1/evaluation_val/mwsc/options-or/results.json
similarity index 100%
rename from evaluation_val/mwsc/options-or/results.json
rename to evaluation_bloomz-7b1/evaluation_val/mwsc/options-or/results.json
diff --git a/evaluation_val/mwsc/what-think/results.json b/evaluation_bloomz-7b1/evaluation_val/mwsc/what-think/results.json
similarity index 100%
rename from evaluation_val/mwsc/what-think/results.json
rename to evaluation_bloomz-7b1/evaluation_val/mwsc/what-think/results.json
diff --git a/evaluation_val/onestop_english/ara_context/results.json b/evaluation_bloomz-7b1/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-7b1/evaluation_val/onestop_english/ara_context/results.json
diff --git a/evaluation_val/onestop_english/assess/results.json b/evaluation_bloomz-7b1/evaluation_val/onestop_english/assess/results.json
similarity index 100%
rename from evaluation_val/onestop_english/assess/results.json
rename to evaluation_bloomz-7b1/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-7b1/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-7b1/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-7b1/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-7b1/evaluation_val/onestop_english/esl_context/results.json
diff --git a/evaluation_val/onestop_english/esl_variation/results.json b/evaluation_bloomz-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/evaluation_val/riddle_sense/most_suitable_answer/results.json
diff --git a/evaluation_val/riddle_sense/question_answering/results.json b/evaluation_bloomz-7b1/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-7b1/evaluation_val/riddle_sense/question_answering/results.json
diff --git a/evaluation_val/riddle_sense/question_to_answer_index/results.json b/evaluation_bloomz-7b1/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-7b1/evaluation_val/riddle_sense/question_to_answer_index/results.json
diff --git a/evaluation_val/scicite/Classify_intent/results.json b/evaluation_bloomz-7b1/evaluation_val/scicite/Classify_intent/results.json
similarity index 100%
rename from evaluation_val/scicite/Classify_intent/results.json
rename to evaluation_bloomz-7b1/evaluation_val/scicite/Classify_intent/results.json
diff --git a/evaluation_val/scicite/Classify_intent_(choices_first)/results.json b/evaluation_bloomz-7b1/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-7b1/evaluation_val/scicite/Classify_intent_(choices_first)/results.json
diff --git a/evaluation_val/scicite/Classify_intent_(select_choice)/results.json b/evaluation_bloomz-7b1/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-7b1/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-7b1/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-7b1/evaluation_val/scicite/Classify_intent_w_section_(select_choice)/results.json
diff --git a/evaluation_val/scicite/can_describe/results.json b/evaluation_bloomz-7b1/evaluation_val/scicite/can_describe/results.json
similarity index 100%
rename from evaluation_val/scicite/can_describe/results.json
rename to evaluation_bloomz-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/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-7b1/evaluation_val/snips_built_in_intents/voice_intent/results.json