Add

4b284b17bc4seed2/evaluation/generation/agg.4b284b17bc4seed2_e2e_nlg_cleaned_generate_text_restaurant_0.json

@@ -0,0 +1 @@
+{"results": [{"task_name": "e2e_nlg_cleaned", "prompt_name": "generate_text_restaurant", "bleu": 1.3562210829072388, "fixed_answer_choice_list": null, "dataset_path": "e2e_nlg_cleaned", "dataset_name": null, "subset": null, "prompt_id": "1acabbc3-c9b9-4624-a684-29faeccff46f", "prompt_jinja": "Given the following data about a restaurant:\n{% for feature in meaning_representation.split(\"]\") %} {% set key = feature.split(\"[\")[0].replace(\",\",\"\") %} {% set value = feature.replace(\",\",\"\").replace(key+\"[\", '''') %}\n{% if value != \"\" %} {{key}} : {{value}} {% endif %}\n{%- endfor %}\nGenerate some text about this restaurant. ||| {{human_reference}}", "prompt_original_task": true, "comment": "", "bleu_stderr": 0.06363294711119734}, {"task_name": "e2e_nlg_cleaned", "prompt_name": "generate_text_restaurant", "rouge1_precision": 0.11009512311383156, "fixed_answer_choice_list": null, "dataset_path": "e2e_nlg_cleaned", "dataset_name": null, "subset": null, "prompt_id": "1acabbc3-c9b9-4624-a684-29faeccff46f", "prompt_jinja": "Given the following data about a restaurant:\n{% for feature in meaning_representation.split(\"]\") %} {% set key = feature.split(\"[\")[0].replace(\",\",\"\") %} {% set value = feature.replace(\",\",\"\").replace(key+\"[\", '''') %}\n{% if value != \"\" %} {{key}} : {{value}} {% endif %}\n{%- endfor %}\nGenerate some text about this restaurant. ||| {{human_reference}}", "prompt_original_task": true, "comment": "", "rouge1_precision_stderr": 0.0020727913604168193}, {"task_name": "e2e_nlg_cleaned", "prompt_name": "generate_text_restaurant", "rouge1_recall": 0.16243977504759669, "fixed_answer_choice_list": null, "dataset_path": "e2e_nlg_cleaned", "dataset_name": null, "subset": null, "prompt_id": "1acabbc3-c9b9-4624-a684-29faeccff46f", "prompt_jinja": "Given the following data about a restaurant:\n{% for feature in meaning_representation.split(\"]\") %} {% set key = feature.split(\"[\")[0].replace(\",\",\"\") %} {% set value = feature.replace(\",\",\"\").replace(key+\"[\", '''') %}\n{% if value != \"\" %} {{key}} : {{value}} {% endif %}\n{%- endfor %}\nGenerate some text about this restaurant. ||| {{human_reference}}", "prompt_original_task": true, "comment": "", "rouge1_recall_stderr": 0.002228768903465126}, {"task_name": "e2e_nlg_cleaned", "prompt_name": "generate_text_restaurant", "rouge1_fmeasure": 0.12013040778127415, "fixed_answer_choice_list": null, "dataset_path": "e2e_nlg_cleaned", "dataset_name": null, "subset": null, "prompt_id": "1acabbc3-c9b9-4624-a684-29faeccff46f", "prompt_jinja": "Given the following data about a restaurant:\n{% for feature in meaning_representation.split(\"]\") %} {% set key = feature.split(\"[\")[0].replace(\",\",\"\") %} {% set value = feature.replace(\",\",\"\").replace(key+\"[\", '''') %}\n{% if value != \"\" %} {{key}} : {{value}} {% endif %}\n{%- endfor %}\nGenerate some text about this restaurant. ||| {{human_reference}}", "prompt_original_task": true, "comment": "", "rouge1_fmeasure_stderr": 0.0017746386859163203}, {"task_name": "e2e_nlg_cleaned", "prompt_name": "generate_text_restaurant", "rouge2_precision": 0.016237573503050346, "fixed_answer_choice_list": null, "dataset_path": "e2e_nlg_cleaned", "dataset_name": null, "subset": null, "prompt_id": "1acabbc3-c9b9-4624-a684-29faeccff46f", "prompt_jinja": "Given the following data about a restaurant:\n{% for feature in meaning_representation.split(\"]\") %} {% set key = feature.split(\"[\")[0].replace(\",\",\"\") %} {% set value = feature.replace(\",\",\"\").replace(key+\"[\", '''') %}\n{% if value != \"\" %} {{key}} : {{value}} {% endif %}\n{%- endfor %}\nGenerate some text about this restaurant. ||| {{human_reference}}", "prompt_original_task": true, "comment": "", "rouge2_precision_stderr": 0.0005791686929917277}, {"task_name": "e2e_nlg_cleaned", "prompt_name": "generate_text_restaurant", "rouge2_recall": 0.027893395046500822, "fixed_answer_choice_list": null, "dataset_path": "e2e_nlg_cleaned", "dataset_name": null, "subset": null, "prompt_id": "1acabbc3-c9b9-4624-a684-29faeccff46f", "prompt_jinja": "Given the following data about a restaurant:\n{% for feature in meaning_representation.split(\"]\") %} {% set key = feature.split(\"[\")[0].replace(\",\",\"\") %} {% set value = feature.replace(\",\",\"\").replace(key+\"[\", '''') %}\n{% if value != \"\" %} {{key}} : {{value}} {% endif %}\n{%- endfor %}\nGenerate some text about this restaurant. ||| {{human_reference}}", "prompt_original_task": true, "comment": "", "rouge2_recall_stderr": 0.001039576769487543}, {"task_name": "e2e_nlg_cleaned", "prompt_name": "generate_text_restaurant", "rouge2_fmeasure": 0.01952192502751282, "fixed_answer_choice_list": null, "dataset_path": "e2e_nlg_cleaned", "dataset_name": null, "subset": null, "prompt_id": "1acabbc3-c9b9-4624-a684-29faeccff46f", "prompt_jinja": "Given the following data about a restaurant:\n{% for feature in meaning_representation.split(\"]\") %} {% set key = feature.split(\"[\")[0].replace(\",\",\"\") %} {% set value = feature.replace(\",\",\"\").replace(key+\"[\", '''') %}\n{% if value != \"\" %} {{key}} : {{value}} {% endif %}\n{%- endfor %}\nGenerate some text about this restaurant. ||| {{human_reference}}", "prompt_original_task": true, "comment": "", "rouge2_fmeasure_stderr": 0.0007045263606140751}, {"task_name": "e2e_nlg_cleaned", "prompt_name": "generate_text_restaurant", "rougeL_precision": 0.10534269523932902, "fixed_answer_choice_list": null, "dataset_path": "e2e_nlg_cleaned", "dataset_name": null, "subset": null, "prompt_id": "1acabbc3-c9b9-4624-a684-29faeccff46f", "prompt_jinja": "Given the following data about a restaurant:\n{% for feature in meaning_representation.split(\"]\") %} {% set key = feature.split(\"[\")[0].replace(\",\",\"\") %} {% set value = feature.replace(\",\",\"\").replace(key+\"[\", '''') %}\n{% if value != \"\" %} {{key}} : {{value}} {% endif %}\n{%- endfor %}\nGenerate some text about this restaurant. ||| {{human_reference}}", "prompt_original_task": true, "comment": "", "rougeL_precision_stderr": 0.0018942650621943632}, {"task_name": "e2e_nlg_cleaned", "prompt_name": "generate_text_restaurant", "rougeL_recall": 0.1589275088375994, "fixed_answer_choice_list": null, "dataset_path": "e2e_nlg_cleaned", "dataset_name": null, "subset": null, "prompt_id": "1acabbc3-c9b9-4624-a684-29faeccff46f", "prompt_jinja": "Given the following data about a restaurant:\n{% for feature in meaning_representation.split(\"]\") %} {% set key = feature.split(\"[\")[0].replace(\",\",\"\") %} {% set value = feature.replace(\",\",\"\").replace(key+\"[\", '''') %}\n{% if value != \"\" %} {{key}} : {{value}} {% endif %}\n{%- endfor %}\nGenerate some text about this restaurant. ||| {{human_reference}}", "prompt_original_task": true, "comment": "", "rougeL_recall_stderr": 0.0021856832752487143}, {"task_name": "e2e_nlg_cleaned", "prompt_name": "generate_text_restaurant", "rougeL_fmeasure": 0.1164630301972042, "fixed_answer_choice_list": null, "dataset_path": "e2e_nlg_cleaned", "dataset_name": null, "subset": null, "prompt_id": "1acabbc3-c9b9-4624-a684-29faeccff46f", "prompt_jinja": "Given the following data about a restaurant:\n{% for feature in meaning_representation.split(\"]\") %} {% set key = feature.split(\"[\")[0].replace(\",\",\"\") %} {% set value = feature.replace(\",\",\"\").replace(key+\"[\", '''') %}\n{% if value != \"\" %} {{key}} : {{value}} {% endif %}\n{%- endfor %}\nGenerate some text about this restaurant. ||| {{human_reference}}", "prompt_original_task": true, "comment": "", "rougeL_fmeasure_stderr": 0.0016922702496737136}, {"task_name": "e2e_nlg_cleaned", "prompt_name": "generate_text_restaurant", "rougeLsum_precision": 0.09138176673759738, "fixed_answer_choice_list": null, "dataset_path": "e2e_nlg_cleaned", "dataset_name": null, "subset": null, "prompt_id": "1acabbc3-c9b9-4624-a684-29faeccff46f", "prompt_jinja": "Given the following data about a restaurant:\n{% for feature in meaning_representation.split(\"]\") %} {% set key = feature.split(\"[\")[0].replace(\",\",\"\") %} {% set value = feature.replace(\",\",\"\").replace(key+\"[\", '''') %}\n{% if value != \"\" %} {{key}} : {{value}} {% endif %}\n{%- endfor %}\nGenerate some text about this restaurant. ||| {{human_reference}}", "prompt_original_task": true, "comment": "", "rougeLsum_precision_stderr": 0.0016976459528528838}, {"task_name": "e2e_nlg_cleaned", "prompt_name": "generate_text_restaurant", "rougeLsum_recall": 0.13851750677859065, "fixed_answer_choice_list": null, "dataset_path": "e2e_nlg_cleaned", "dataset_name": null, "subset": null, "prompt_id": "1acabbc3-c9b9-4624-a684-29faeccff46f", "prompt_jinja": "Given the following data about a restaurant:\n{% for feature in meaning_representation.split(\"]\") %} {% set key = feature.split(\"[\")[0].replace(\",\",\"\") %} {% set value = feature.replace(\",\",\"\").replace(key+\"[\", '''') %}\n{% if value != \"\" %} {{key}} : {{value}} {% endif %}\n{%- endfor %}\nGenerate some text about this restaurant. ||| {{human_reference}}", "prompt_original_task": true, "comment": "", "rougeLsum_recall_stderr": 0.001981721576730165}, {"task_name": "e2e_nlg_cleaned", "prompt_name": "generate_text_restaurant", "rougeLsum_fmeasure": 0.1010326136243227, "fixed_answer_choice_list": null, "dataset_path": "e2e_nlg_cleaned", "dataset_name": null, "subset": null, "prompt_id": "1acabbc3-c9b9-4624-a684-29faeccff46f", "prompt_jinja": "Given the following data about a restaurant:\n{% for feature in meaning_representation.split(\"]\") %} {% set key = feature.split(\"[\")[0].replace(\",\",\"\") %} {% set value = feature.replace(\",\",\"\").replace(key+\"[\", '''') %}\n{% if value != \"\" %} {{key}} : {{value}} {% endif %}\n{%- endfor %}\nGenerate some text about this restaurant. ||| {{human_reference}}", "prompt_original_task": true, "comment": "", "rougeLsum_fmeasure_stderr": 0.0015162989057978518}], "config": {"model": "hf-causal", "model_args": "pretrained=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4seeds/4b284b17bc4seed2/transformers,use_accelerate=True,tokenizer=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/gpt2,dtype=bfloat16", "task_args": "", "num_fewshot": 0, "batch_size": 16, "device": "cuda", "use_cache": false, "limit": 3000, "bootstrap_iters": 10, "seed": 1234}}

4b284b17bc4seed2/evaluation/generation/agg.4b284b17bc4seed2_gem_xsum_article_DOC_summary_0.json

@@ -0,0 +1 @@
+{"results": [{"task_name": "gem_xsum", "prompt_name": "article_DOC_summary", "rouge1_precision": 0.16625735855106855, "fixed_answer_choice_list": null, "dataset_path": "GEM/xsum", "dataset_name": null, "subset": "", "prompt_id": "a8d4ecfa-c944-44d5-878c-04fd5db59e64", "prompt_jinja": "Article: {{document}}\n\nSummary: ||| {{target}}", "prompt_original_task": true, "comment": "", "rouge1_precision_stderr": 0.0024463651813182914}, {"task_name": "gem_xsum", "prompt_name": "article_DOC_summary", "rouge1_recall": 0.3230884270864929, "fixed_answer_choice_list": null, "dataset_path": "GEM/xsum", "dataset_name": null, "subset": "", "prompt_id": "a8d4ecfa-c944-44d5-878c-04fd5db59e64", "prompt_jinja": "Article: {{document}}\n\nSummary: ||| {{target}}", "prompt_original_task": true, "comment": "", "rouge1_recall_stderr": 0.004359311826645138}, {"task_name": "gem_xsum", "prompt_name": "article_DOC_summary", "rouge1_fmeasure": 0.20872501710774183, "fixed_answer_choice_list": null, "dataset_path": "GEM/xsum", "dataset_name": null, "subset": "", "prompt_id": "a8d4ecfa-c944-44d5-878c-04fd5db59e64", "prompt_jinja": "Article: {{document}}\n\nSummary: ||| {{target}}", "prompt_original_task": true, "comment": "", "rouge1_fmeasure_stderr": 0.0025765264814385574}, {"task_name": "gem_xsum", "prompt_name": "article_DOC_summary", "rouge2_precision": 0.03520681144190837, "fixed_answer_choice_list": null, "dataset_path": "GEM/xsum", "dataset_name": null, "subset": "", "prompt_id": "a8d4ecfa-c944-44d5-878c-04fd5db59e64", "prompt_jinja": "Article: {{document}}\n\nSummary: ||| {{target}}", "prompt_original_task": true, "comment": "", "rouge2_precision_stderr": 0.0013363705236252219}, {"task_name": "gem_xsum", "prompt_name": "article_DOC_summary", "rouge2_recall": 0.07441126206565717, "fixed_answer_choice_list": null, "dataset_path": "GEM/xsum", "dataset_name": null, "subset": "", "prompt_id": "a8d4ecfa-c944-44d5-878c-04fd5db59e64", "prompt_jinja": "Article: {{document}}\n\nSummary: ||| {{target}}", "prompt_original_task": true, "comment": "", "rouge2_recall_stderr": 0.0027685648685041907}, {"task_name": "gem_xsum", "prompt_name": "article_DOC_summary", "rouge2_fmeasure": 0.04541415352191168, "fixed_answer_choice_list": null, "dataset_path": "GEM/xsum", "dataset_name": null, "subset": "", "prompt_id": "a8d4ecfa-c944-44d5-878c-04fd5db59e64", "prompt_jinja": "Article: {{document}}\n\nSummary: ||| {{target}}", "prompt_original_task": true, "comment": "", "rouge2_fmeasure_stderr": 0.0016019047132235528}, {"task_name": "gem_xsum", "prompt_name": "article_DOC_summary", "rougeL_precision": 0.12527352638980324, "fixed_answer_choice_list": null, "dataset_path": "GEM/xsum", "dataset_name": null, "subset": "", "prompt_id": "a8d4ecfa-c944-44d5-878c-04fd5db59e64", "prompt_jinja": "Article: {{document}}\n\nSummary: ||| {{target}}", "prompt_original_task": true, "comment": "", "rougeL_precision_stderr": 0.0018832520008059408}, {"task_name": "gem_xsum", "prompt_name": "article_DOC_summary", "rougeL_recall": 0.24649639309519972, "fixed_answer_choice_list": null, "dataset_path": "GEM/xsum", "dataset_name": null, "subset": "", "prompt_id": "a8d4ecfa-c944-44d5-878c-04fd5db59e64", "prompt_jinja": "Article: {{document}}\n\nSummary: ||| {{target}}", "prompt_original_task": true, "comment": "", "rougeL_recall_stderr": 0.0035131508278542895}, {"task_name": "gem_xsum", "prompt_name": "article_DOC_summary", "rougeL_fmeasure": 0.15771639364031548, "fixed_answer_choice_list": null, "dataset_path": "GEM/xsum", "dataset_name": null, "subset": "", "prompt_id": "a8d4ecfa-c944-44d5-878c-04fd5db59e64", "prompt_jinja": "Article: {{document}}\n\nSummary: ||| {{target}}", "prompt_original_task": true, "comment": "", "rougeL_fmeasure_stderr": 0.001968467651802171}, {"task_name": "gem_xsum", "prompt_name": "article_DOC_summary", "rougeLsum_precision": 0.12771639136222968, "fixed_answer_choice_list": null, "dataset_path": "GEM/xsum", "dataset_name": null, "subset": "", "prompt_id": "a8d4ecfa-c944-44d5-878c-04fd5db59e64", "prompt_jinja": "Article: {{document}}\n\nSummary: ||| {{target}}", "prompt_original_task": true, "comment": "", "rougeLsum_precision_stderr": 0.001971864037023057}, {"task_name": "gem_xsum", "prompt_name": "article_DOC_summary", "rougeLsum_recall": 0.2517880179492348, "fixed_answer_choice_list": null, "dataset_path": "GEM/xsum", "dataset_name": null, "subset": "", "prompt_id": "a8d4ecfa-c944-44d5-878c-04fd5db59e64", "prompt_jinja": "Article: {{document}}\n\nSummary: ||| {{target}}", "prompt_original_task": true, "comment": "", "rougeLsum_recall_stderr": 0.003783114418327505}, {"task_name": "gem_xsum", "prompt_name": "article_DOC_summary", "rougeLsum_fmeasure": 0.16106544122061872, "fixed_answer_choice_list": null, "dataset_path": "GEM/xsum", "dataset_name": null, "subset": "", "prompt_id": "a8d4ecfa-c944-44d5-878c-04fd5db59e64", "prompt_jinja": "Article: {{document}}\n\nSummary: ||| {{target}}", "prompt_original_task": true, "comment": "", "rougeLsum_fmeasure_stderr": 0.002144088469758669}, {"task_name": "gem_xsum", "prompt_name": "article_DOC_summary", "bleu": 2.036660647928376, "fixed_answer_choice_list": null, "dataset_path": "GEM/xsum", "dataset_name": null, "subset": "", "prompt_id": "a8d4ecfa-c944-44d5-878c-04fd5db59e64", "prompt_jinja": "Article: {{document}}\n\nSummary: ||| {{target}}", "prompt_original_task": true, "comment": "", "bleu_stderr": 0.0886671989991428}], "config": {"model": "hf-causal", "model_args": "pretrained=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4seeds/4b284b17bc4seed2/transformers,use_accelerate=True,tokenizer=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/gpt2,dtype=bfloat16", "task_args": "", "num_fewshot": 0, "batch_size": 16, "device": "cuda", "use_cache": false, "limit": 3000, "bootstrap_iters": 10, "seed": 1234}}

4b284b17bc4seed2/evaluation/generation/agg.4b284b17bc4seed2_gem_xsum_article_DOC_summary_1.json

@@ -0,0 +1 @@
+{"results": [{"task_name": "gem_xsum", "prompt_name": "article_DOC_summary", "rouge1_precision": 0.13473651953646212, "fixed_answer_choice_list": null, "dataset_path": "GEM/xsum", "dataset_name": null, "subset": "", "prompt_id": "a8d4ecfa-c944-44d5-878c-04fd5db59e64", "prompt_jinja": "Article: {{document}}\n\nSummary: ||| {{target}}", "prompt_original_task": true, "comment": "", "rouge1_precision_stderr": 0.0018830579225946448}, {"task_name": "gem_xsum", "prompt_name": "article_DOC_summary", "rouge1_recall": 0.33263109371075555, "fixed_answer_choice_list": null, "dataset_path": "GEM/xsum", "dataset_name": null, "subset": "", "prompt_id": "a8d4ecfa-c944-44d5-878c-04fd5db59e64", "prompt_jinja": "Article: {{document}}\n\nSummary: ||| {{target}}", "prompt_original_task": true, "comment": "", "rouge1_recall_stderr": 0.0043737298284748935}, {"task_name": "gem_xsum", "prompt_name": "article_DOC_summary", "rouge1_fmeasure": 0.18955906449065837, "fixed_answer_choice_list": null, "dataset_path": "GEM/xsum", "dataset_name": null, "subset": "", "prompt_id": "a8d4ecfa-c944-44d5-878c-04fd5db59e64", "prompt_jinja": "Article: {{document}}\n\nSummary: ||| {{target}}", "prompt_original_task": true, "comment": "", "rouge1_fmeasure_stderr": 0.0025437030392446553}, {"task_name": "gem_xsum", "prompt_name": "article_DOC_summary", "rouge2_precision": 0.02979980903695056, "fixed_answer_choice_list": null, "dataset_path": "GEM/xsum", "dataset_name": null, "subset": "", "prompt_id": "a8d4ecfa-c944-44d5-878c-04fd5db59e64", "prompt_jinja": "Article: {{document}}\n\nSummary: ||| {{target}}", "prompt_original_task": true, "comment": "", "rouge2_precision_stderr": 0.0010699057640840117}, {"task_name": "gem_xsum", "prompt_name": "article_DOC_summary", "rouge2_recall": 0.07661622521673665, "fixed_answer_choice_list": null, "dataset_path": "GEM/xsum", "dataset_name": null, "subset": "", "prompt_id": "a8d4ecfa-c944-44d5-878c-04fd5db59e64", "prompt_jinja": "Article: {{document}}\n\nSummary: ||| {{target}}", "prompt_original_task": true, "comment": "", "rouge2_recall_stderr": 0.00277147224226074}, {"task_name": "gem_xsum", "prompt_name": "article_DOC_summary", "rouge2_fmeasure": 0.04240941625507146, "fixed_answer_choice_list": null, "dataset_path": "GEM/xsum", "dataset_name": null, "subset": "", "prompt_id": "a8d4ecfa-c944-44d5-878c-04fd5db59e64", "prompt_jinja": "Article: {{document}}\n\nSummary: ||| {{target}}", "prompt_original_task": true, "comment": "", "rouge2_fmeasure_stderr": 0.0015134151293483083}, {"task_name": "gem_xsum", "prompt_name": "article_DOC_summary", "rougeL_precision": 0.10545507286869261, "fixed_answer_choice_list": null, "dataset_path": "GEM/xsum", "dataset_name": null, "subset": "", "prompt_id": "a8d4ecfa-c944-44d5-878c-04fd5db59e64", "prompt_jinja": "Article: {{document}}\n\nSummary: ||| {{target}}", "prompt_original_task": true, "comment": "", "rougeL_precision_stderr": 0.0014282925798028265}, {"task_name": "gem_xsum", "prompt_name": "article_DOC_summary", "rougeL_recall": 0.2622174946749768, "fixed_answer_choice_list": null, "dataset_path": "GEM/xsum", "dataset_name": null, "subset": "", "prompt_id": "a8d4ecfa-c944-44d5-878c-04fd5db59e64", "prompt_jinja": "Article: {{document}}\n\nSummary: ||| {{target}}", "prompt_original_task": true, "comment": "", "rougeL_recall_stderr": 0.0034532690725241335}, {"task_name": "gem_xsum", "prompt_name": "article_DOC_summary", "rougeL_fmeasure": 0.14860417887228614, "fixed_answer_choice_list": null, "dataset_path": "GEM/xsum", "dataset_name": null, "subset": "", "prompt_id": "a8d4ecfa-c944-44d5-878c-04fd5db59e64", "prompt_jinja": "Article: {{document}}\n\nSummary: ||| {{target}}", "prompt_original_task": true, "comment": "", "rougeL_fmeasure_stderr": 0.0019427079439186818}, {"task_name": "gem_xsum", "prompt_name": "article_DOC_summary", "rougeLsum_precision": 0.10644896438074114, "fixed_answer_choice_list": null, "dataset_path": "GEM/xsum", "dataset_name": null, "subset": "", "prompt_id": "a8d4ecfa-c944-44d5-878c-04fd5db59e64", "prompt_jinja": "Article: {{document}}\n\nSummary: ||| {{target}}", "prompt_original_task": true, "comment": "", "rougeLsum_precision_stderr": 0.0015535641404371687}, {"task_name": "gem_xsum", "prompt_name": "article_DOC_summary", "rougeLsum_recall": 0.26570822729983273, "fixed_answer_choice_list": null, "dataset_path": "GEM/xsum", "dataset_name": null, "subset": "", "prompt_id": "a8d4ecfa-c944-44d5-878c-04fd5db59e64", "prompt_jinja": "Article: {{document}}\n\nSummary: ||| {{target}}", "prompt_original_task": true, "comment": "", "rougeLsum_recall_stderr": 0.003832781853814249}, {"task_name": "gem_xsum", "prompt_name": "article_DOC_summary", "rougeLsum_fmeasure": 0.1502154921172796, "fixed_answer_choice_list": null, "dataset_path": "GEM/xsum", "dataset_name": null, "subset": "", "prompt_id": "a8d4ecfa-c944-44d5-878c-04fd5db59e64", "prompt_jinja": "Article: {{document}}\n\nSummary: ||| {{target}}", "prompt_original_task": true, "comment": "", "rougeLsum_fmeasure_stderr": 0.0021382854049990575}, {"task_name": "gem_xsum", "prompt_name": "article_DOC_summary", "bleu": 1.6913128558245667, "fixed_answer_choice_list": null, "dataset_path": "GEM/xsum", "dataset_name": null, "subset": "", "prompt_id": "a8d4ecfa-c944-44d5-878c-04fd5db59e64", "prompt_jinja": "Article: {{document}}\n\nSummary: ||| {{target}}", "prompt_original_task": true, "comment": "", "bleu_stderr": 0.09521771990115542}], "config": {"model": "hf-causal", "model_args": "pretrained=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4seeds/4b284b17bc4seed2/transformers,use_accelerate=True,tokenizer=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/gpt2,dtype=bfloat16", "task_args": "", "num_fewshot": 1, "batch_size": 16, "device": "cuda", "use_cache": false, "limit": 3000, "bootstrap_iters": 10, "seed": 1234}}

4b284b17bc4seed2/evaluation/generation/examples.4b284b17bc4seed2_e2e_nlg_cleaned_generate_text_restaurant_0.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5acf9e32092153cf8f0af5c91853b34d9b865cdebba19399b5ad1f802f36bb55
+size 4454140

4b284b17bc4seed2/evaluation/generation/examples.4b284b17bc4seed2_gem_xsum_article_DOC_summary_0.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ed45a900f08a095774f6f224b39bd55c362922ea32bc8f0c0bd18cef7b82e73a
+size 2788220

4b284b17bc4seed2/evaluation/generation/examples.4b284b17bc4seed2_gem_xsum_article_DOC_summary_1.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:81016c90af5ffef779f0cb474c304822fd5faf4d585ef531961dff3876ada5a4
+size 5106203

4b284b17bc4seed2/evaluation/generation/slim.4b284b17bc4seed2_e2e_nlg_cleaned_generate_text_restaurant_0.json

@@ -0,0 +1,133 @@
+{
+  "results": [
+    {
+      "task_name": "e2e_nlg_cleaned",
+      "prompt_name": "generate_text_restaurant",
+      "bleu": 1.3562210829072388,
+      "dataset_path": "e2e_nlg_cleaned",
+      "dataset_name": null,
+      "subset": null,
+      "bleu_stderr": 0.06363294711119734
+    },
+    {
+      "task_name": "e2e_nlg_cleaned",
+      "prompt_name": "generate_text_restaurant",
+      "rouge1_precision": 0.11009512311383156,
+      "dataset_path": "e2e_nlg_cleaned",
+      "dataset_name": null,
+      "subset": null,
+      "rouge1_precision_stderr": 0.0020727913604168193
+    },
+    {
+      "task_name": "e2e_nlg_cleaned",
+      "prompt_name": "generate_text_restaurant",
+      "rouge1_recall": 0.16243977504759669,
+      "dataset_path": "e2e_nlg_cleaned",
+      "dataset_name": null,
+      "subset": null,
+      "rouge1_recall_stderr": 0.002228768903465126
+    },
+    {
+      "task_name": "e2e_nlg_cleaned",
+      "prompt_name": "generate_text_restaurant",
+      "rouge1_fmeasure": 0.12013040778127415,
+      "dataset_path": "e2e_nlg_cleaned",
+      "dataset_name": null,
+      "subset": null,
+      "rouge1_fmeasure_stderr": 0.0017746386859163203
+    },
+    {
+      "task_name": "e2e_nlg_cleaned",
+      "prompt_name": "generate_text_restaurant",
+      "rouge2_precision": 0.016237573503050346,
+      "dataset_path": "e2e_nlg_cleaned",
+      "dataset_name": null,
+      "subset": null,
+      "rouge2_precision_stderr": 0.0005791686929917277
+    },
+    {
+      "task_name": "e2e_nlg_cleaned",
+      "prompt_name": "generate_text_restaurant",
+      "rouge2_recall": 0.027893395046500822,
+      "dataset_path": "e2e_nlg_cleaned",
+      "dataset_name": null,
+      "subset": null,
+      "rouge2_recall_stderr": 0.001039576769487543
+    },
+    {
+      "task_name": "e2e_nlg_cleaned",
+      "prompt_name": "generate_text_restaurant",
+      "rouge2_fmeasure": 0.01952192502751282,
+      "dataset_path": "e2e_nlg_cleaned",
+      "dataset_name": null,
+      "subset": null,
+      "rouge2_fmeasure_stderr": 0.0007045263606140751
+    },
+    {
+      "task_name": "e2e_nlg_cleaned",
+      "prompt_name": "generate_text_restaurant",
+      "rougeL_precision": 0.10534269523932902,
+      "dataset_path": "e2e_nlg_cleaned",
+      "dataset_name": null,
+      "subset": null,
+      "rougeL_precision_stderr": 0.0018942650621943632
+    },
+    {
+      "task_name": "e2e_nlg_cleaned",
+      "prompt_name": "generate_text_restaurant",
+      "rougeL_recall": 0.1589275088375994,
+      "dataset_path": "e2e_nlg_cleaned",
+      "dataset_name": null,
+      "subset": null,
+      "rougeL_recall_stderr": 0.0021856832752487143
+    },
+    {
+      "task_name": "e2e_nlg_cleaned",
+      "prompt_name": "generate_text_restaurant",
+      "rougeL_fmeasure": 0.1164630301972042,
+      "dataset_path": "e2e_nlg_cleaned",
+      "dataset_name": null,
+      "subset": null,
+      "rougeL_fmeasure_stderr": 0.0016922702496737136
+    },
+    {
+      "task_name": "e2e_nlg_cleaned",
+      "prompt_name": "generate_text_restaurant",
+      "rougeLsum_precision": 0.09138176673759738,
+      "dataset_path": "e2e_nlg_cleaned",
+      "dataset_name": null,
+      "subset": null,
+      "rougeLsum_precision_stderr": 0.0016976459528528838
+    },
+    {
+      "task_name": "e2e_nlg_cleaned",
+      "prompt_name": "generate_text_restaurant",
+      "rougeLsum_recall": 0.13851750677859065,
+      "dataset_path": "e2e_nlg_cleaned",
+      "dataset_name": null,
+      "subset": null,
+      "rougeLsum_recall_stderr": 0.001981721576730165
+    },
+    {
+      "task_name": "e2e_nlg_cleaned",
+      "prompt_name": "generate_text_restaurant",
+      "rougeLsum_fmeasure": 0.1010326136243227,
+      "dataset_path": "e2e_nlg_cleaned",
+      "dataset_name": null,
+      "subset": null,
+      "rougeLsum_fmeasure_stderr": 0.0015162989057978518
+    }
+  ],
+  "config": {
+    "model": "hf-causal",
+    "model_args": "pretrained=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4seeds/4b284b17bc4seed2/transformers,use_accelerate=True,tokenizer=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/gpt2,dtype=bfloat16",
+    "task_args": "",
+    "num_fewshot": 0,
+    "batch_size": 16,
+    "device": "cuda",
+    "use_cache": false,
+    "limit": 3000,
+    "bootstrap_iters": 10,
+    "seed": 1234
+  }
+}

4b284b17bc4seed2/evaluation/generation/slim.4b284b17bc4seed2_gem_xsum_article_DOC_summary_0.json

@@ -0,0 +1,133 @@
+{
+  "results": [
+    {
+      "task_name": "gem_xsum",
+      "prompt_name": "article_DOC_summary",
+      "rouge1_precision": 0.16625735855106855,
+      "dataset_path": "GEM/xsum",
+      "dataset_name": null,
+      "subset": "",
+      "rouge1_precision_stderr": 0.0024463651813182914
+    },
+    {
+      "task_name": "gem_xsum",
+      "prompt_name": "article_DOC_summary",
+      "rouge1_recall": 0.3230884270864929,
+      "dataset_path": "GEM/xsum",
+      "dataset_name": null,
+      "subset": "",
+      "rouge1_recall_stderr": 0.004359311826645138
+    },
+    {
+      "task_name": "gem_xsum",
+      "prompt_name": "article_DOC_summary",
+      "rouge1_fmeasure": 0.20872501710774183,
+      "dataset_path": "GEM/xsum",
+      "dataset_name": null,
+      "subset": "",
+      "rouge1_fmeasure_stderr": 0.0025765264814385574
+    },
+    {
+      "task_name": "gem_xsum",
+      "prompt_name": "article_DOC_summary",
+      "rouge2_precision": 0.03520681144190837,
+      "dataset_path": "GEM/xsum",
+      "dataset_name": null,
+      "subset": "",
+      "rouge2_precision_stderr": 0.0013363705236252219
+    },
+    {
+      "task_name": "gem_xsum",
+      "prompt_name": "article_DOC_summary",
+      "rouge2_recall": 0.07441126206565717,
+      "dataset_path": "GEM/xsum",
+      "dataset_name": null,
+      "subset": "",
+      "rouge2_recall_stderr": 0.0027685648685041907
+    },
+    {
+      "task_name": "gem_xsum",
+      "prompt_name": "article_DOC_summary",
+      "rouge2_fmeasure": 0.04541415352191168,
+      "dataset_path": "GEM/xsum",
+      "dataset_name": null,
+      "subset": "",
+      "rouge2_fmeasure_stderr": 0.0016019047132235528
+    },
+    {
+      "task_name": "gem_xsum",
+      "prompt_name": "article_DOC_summary",
+      "rougeL_precision": 0.12527352638980324,
+      "dataset_path": "GEM/xsum",
+      "dataset_name": null,
+      "subset": "",
+      "rougeL_precision_stderr": 0.0018832520008059408
+    },
+    {
+      "task_name": "gem_xsum",
+      "prompt_name": "article_DOC_summary",
+      "rougeL_recall": 0.24649639309519972,
+      "dataset_path": "GEM/xsum",
+      "dataset_name": null,
+      "subset": "",
+      "rougeL_recall_stderr": 0.0035131508278542895
+    },
+    {
+      "task_name": "gem_xsum",
+      "prompt_name": "article_DOC_summary",
+      "rougeL_fmeasure": 0.15771639364031548,
+      "dataset_path": "GEM/xsum",
+      "dataset_name": null,
+      "subset": "",
+      "rougeL_fmeasure_stderr": 0.001968467651802171
+    },
+    {
+      "task_name": "gem_xsum",
+      "prompt_name": "article_DOC_summary",
+      "rougeLsum_precision": 0.12771639136222968,
+      "dataset_path": "GEM/xsum",
+      "dataset_name": null,
+      "subset": "",
+      "rougeLsum_precision_stderr": 0.001971864037023057
+    },
+    {
+      "task_name": "gem_xsum",
+      "prompt_name": "article_DOC_summary",
+      "rougeLsum_recall": 0.2517880179492348,
+      "dataset_path": "GEM/xsum",
+      "dataset_name": null,
+      "subset": "",
+      "rougeLsum_recall_stderr": 0.003783114418327505
+    },
+    {
+      "task_name": "gem_xsum",
+      "prompt_name": "article_DOC_summary",
+      "rougeLsum_fmeasure": 0.16106544122061872,
+      "dataset_path": "GEM/xsum",
+      "dataset_name": null,
+      "subset": "",
+      "rougeLsum_fmeasure_stderr": 0.002144088469758669
+    },
+    {
+      "task_name": "gem_xsum",
+      "prompt_name": "article_DOC_summary",
+      "bleu": 2.036660647928376,
+      "dataset_path": "GEM/xsum",
+      "dataset_name": null,
+      "subset": "",
+      "bleu_stderr": 0.0886671989991428
+    }
+  ],
+  "config": {
+    "model": "hf-causal",
+    "model_args": "pretrained=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4seeds/4b284b17bc4seed2/transformers,use_accelerate=True,tokenizer=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/gpt2,dtype=bfloat16",
+    "task_args": "",
+    "num_fewshot": 0,
+    "batch_size": 16,
+    "device": "cuda",
+    "use_cache": false,
+    "limit": 3000,
+    "bootstrap_iters": 10,
+    "seed": 1234
+  }
+}

4b284b17bc4seed2/evaluation/generation/slim.4b284b17bc4seed2_gem_xsum_article_DOC_summary_1.json

@@ -0,0 +1,133 @@
+{
+  "results": [
+    {
+      "task_name": "gem_xsum",
+      "prompt_name": "article_DOC_summary",
+      "rouge1_precision": 0.13473651953646212,
+      "dataset_path": "GEM/xsum",
+      "dataset_name": null,
+      "subset": "",
+      "rouge1_precision_stderr": 0.0018830579225946448
+    },
+    {
+      "task_name": "gem_xsum",
+      "prompt_name": "article_DOC_summary",
+      "rouge1_recall": 0.33263109371075555,
+      "dataset_path": "GEM/xsum",
+      "dataset_name": null,
+      "subset": "",
+      "rouge1_recall_stderr": 0.0043737298284748935
+    },
+    {
+      "task_name": "gem_xsum",
+      "prompt_name": "article_DOC_summary",
+      "rouge1_fmeasure": 0.18955906449065837,
+      "dataset_path": "GEM/xsum",
+      "dataset_name": null,
+      "subset": "",
+      "rouge1_fmeasure_stderr": 0.0025437030392446553
+    },
+    {
+      "task_name": "gem_xsum",
+      "prompt_name": "article_DOC_summary",
+      "rouge2_precision": 0.02979980903695056,
+      "dataset_path": "GEM/xsum",
+      "dataset_name": null,
+      "subset": "",
+      "rouge2_precision_stderr": 0.0010699057640840117
+    },
+    {
+      "task_name": "gem_xsum",
+      "prompt_name": "article_DOC_summary",
+      "rouge2_recall": 0.07661622521673665,
+      "dataset_path": "GEM/xsum",
+      "dataset_name": null,
+      "subset": "",
+      "rouge2_recall_stderr": 0.00277147224226074
+    },
+    {
+      "task_name": "gem_xsum",
+      "prompt_name": "article_DOC_summary",
+      "rouge2_fmeasure": 0.04240941625507146,
+      "dataset_path": "GEM/xsum",
+      "dataset_name": null,
+      "subset": "",
+      "rouge2_fmeasure_stderr": 0.0015134151293483083
+    },
+    {
+      "task_name": "gem_xsum",
+      "prompt_name": "article_DOC_summary",
+      "rougeL_precision": 0.10545507286869261,
+      "dataset_path": "GEM/xsum",
+      "dataset_name": null,
+      "subset": "",
+      "rougeL_precision_stderr": 0.0014282925798028265
+    },
+    {
+      "task_name": "gem_xsum",
+      "prompt_name": "article_DOC_summary",
+      "rougeL_recall": 0.2622174946749768,
+      "dataset_path": "GEM/xsum",
+      "dataset_name": null,
+      "subset": "",
+      "rougeL_recall_stderr": 0.0034532690725241335
+    },
+    {
+      "task_name": "gem_xsum",
+      "prompt_name": "article_DOC_summary",
+      "rougeL_fmeasure": 0.14860417887228614,
+      "dataset_path": "GEM/xsum",
+      "dataset_name": null,
+      "subset": "",
+      "rougeL_fmeasure_stderr": 0.0019427079439186818
+    },
+    {
+      "task_name": "gem_xsum",
+      "prompt_name": "article_DOC_summary",
+      "rougeLsum_precision": 0.10644896438074114,
+      "dataset_path": "GEM/xsum",
+      "dataset_name": null,
+      "subset": "",
+      "rougeLsum_precision_stderr": 0.0015535641404371687
+    },
+    {
+      "task_name": "gem_xsum",
+      "prompt_name": "article_DOC_summary",
+      "rougeLsum_recall": 0.26570822729983273,
+      "dataset_path": "GEM/xsum",
+      "dataset_name": null,
+      "subset": "",
+      "rougeLsum_recall_stderr": 0.003832781853814249
+    },
+    {
+      "task_name": "gem_xsum",
+      "prompt_name": "article_DOC_summary",
+      "rougeLsum_fmeasure": 0.1502154921172796,
+      "dataset_path": "GEM/xsum",
+      "dataset_name": null,
+      "subset": "",
+      "rougeLsum_fmeasure_stderr": 0.0021382854049990575
+    },
+    {
+      "task_name": "gem_xsum",
+      "prompt_name": "article_DOC_summary",
+      "bleu": 1.6913128558245667,
+      "dataset_path": "GEM/xsum",
+      "dataset_name": null,
+      "subset": "",
+      "bleu_stderr": 0.09521771990115542
+    }
+  ],
+  "config": {
+    "model": "hf-causal",
+    "model_args": "pretrained=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4seeds/4b284b17bc4seed2/transformers,use_accelerate=True,tokenizer=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/gpt2,dtype=bfloat16",
+    "task_args": "",
+    "num_fewshot": 1,
+    "batch_size": 16,
+    "device": "cuda",
+    "use_cache": false,
+    "limit": 3000,
+    "bootstrap_iters": 10,
+    "seed": 1234
+  }
+}

4b284b17bc4seed2/evaluation/rankeval/4b284b17bc4seed2_0.json

@@ -0,0 +1,32 @@
+{
+    "results": {
+        "anli_r1": {
+            "acc": 0.332,
+            "acc_stderr": 0.014899597242811482
+        },
+        "anli_r2": {
+            "acc": 0.329,
+            "acc_stderr": 0.014865395385928366
+        },
+        "anli_r3": {
+            "acc": 0.34,
+            "acc_stderr": 0.013680495725767789
+        },
+        "cb": {
+            "acc": 0.39285714285714285,
+            "acc_stderr": 0.0658538889806635,
+            "f1": 0.20817204301075268
+        },
+        "copa": {
+            "acc": 0.78,
+            "acc_stderr": 0.04163331998932261
+        }
+    },
+    "versions": {
+        "anli_r1": 0,
+        "anli_r2": 0,
+        "anli_r3": 0,
+        "cb": 1,
+        "copa": 0
+    }
+}

4b284b17bc4seed2/evaluation/rankeval/4b284b17bc4seed2_0_lm-eval_global_step80108_2023-02-26-09-23-54_0shots_backup.json

@@ -0,0 +1,32 @@
+{
+    "results": {
+        "anli_r1": {
+            "acc": 0.332,
+            "acc_stderr": 0.014899597242811482
+        },
+        "anli_r2": {
+            "acc": 0.329,
+            "acc_stderr": 0.014865395385928366
+        },
+        "anli_r3": {
+            "acc": 0.34,
+            "acc_stderr": 0.013680495725767789
+        },
+        "cb": {
+            "acc": 0.39285714285714285,
+            "acc_stderr": 0.0658538889806635,
+            "f1": 0.20817204301075268
+        },
+        "copa": {
+            "acc": 0.78,
+            "acc_stderr": 0.04163331998932261
+        }
+    },
+    "versions": {
+        "anli_r1": 0,
+        "anli_r2": 0,
+        "anli_r3": 0,
+        "cb": 1,
+        "copa": 0
+    }
+}

4b284b17bc4seed2/evaluation/rankeval/4b284b17bc4seed2_1.json

@@ -0,0 +1,32 @@
+{
+    "results": {
+        "anli_r1": {
+            "acc": 0.31,
+            "acc_stderr": 0.0146326386586329
+        },
+        "anli_r2": {
+            "acc": 0.324,
+            "acc_stderr": 0.014806864733738852
+        },
+        "anli_r3": {
+            "acc": 0.3416666666666667,
+            "acc_stderr": 0.013696658778002517
+        },
+        "cb": {
+            "acc": 0.5,
+            "acc_stderr": 0.06741998624632421,
+            "f1": 0.33520950594121324
+        },
+        "copa": {
+            "acc": 0.76,
+            "acc_stderr": 0.04292346959909283
+        }
+    },
+    "versions": {
+        "anli_r1": 0,
+        "anli_r2": 0,
+        "anli_r3": 0,
+        "cb": 1,
+        "copa": 0
+    }
+}

4b284b17bc4seed2/evaluation/rankeval/4b284b17bc4seed2_1_lm-eval_global_step80108_2023-02-26-09-23-54_1shots_backup.json

@@ -0,0 +1,32 @@
+{
+    "results": {
+        "anli_r1": {
+            "acc": 0.31,
+            "acc_stderr": 0.0146326386586329
+        },
+        "anli_r2": {
+            "acc": 0.324,
+            "acc_stderr": 0.014806864733738852
+        },
+        "anli_r3": {
+            "acc": 0.3416666666666667,
+            "acc_stderr": 0.013696658778002517
+        },
+        "cb": {
+            "acc": 0.5,
+            "acc_stderr": 0.06741998624632421,
+            "f1": 0.33520950594121324
+        },
+        "copa": {
+            "acc": 0.76,
+            "acc_stderr": 0.04292346959909283
+        }
+    },
+    "versions": {
+        "anli_r1": 0,
+        "anli_r2": 0,
+        "anli_r3": 0,
+        "cb": 1,
+        "copa": 0
+    }
+}

4b284b17bc4seed2/evaluation/rankeval/4b284b17bc4seed2_2.json

@@ -0,0 +1,16 @@
+{
+    "results": {
+        "anli_r1": {
+            "acc": 0.34,
+            "acc_stderr": 0.014987482264363937
+        },
+        "anli_r2": {
+            "acc": 0.321,
+            "acc_stderr": 0.014770821817934644
+        }
+    },
+    "versions": {
+        "anli_r1": 0,
+        "anli_r2": 0
+    }
+}

4b284b17bc4seed2/evaluation/rankeval/4b284b17bc4seed2_2_lm-eval_global_step80108_2023-02-26-09-23-54_2shots_backup.json

@@ -0,0 +1,16 @@
+{
+    "results": {
+        "anli_r1": {
+            "acc": 0.34,
+            "acc_stderr": 0.014987482264363937
+        },
+        "anli_r2": {
+            "acc": 0.321,
+            "acc_stderr": 0.014770821817934644
+        }
+    },
+    "versions": {
+        "anli_r1": 0,
+        "anli_r2": 0
+    }
+}

4b284b17bc4seed2/evaluation/rankeval/4b284b17bc4seed2_3.json

@@ -0,0 +1,16 @@
+{
+    "results": {
+        "anli_r1": {
+            "acc": 0.329,
+            "acc_stderr": 0.014865395385928366
+        },
+        "anli_r2": {
+            "acc": 0.327,
+            "acc_stderr": 0.01484221315341125
+        }
+    },
+    "versions": {
+        "anli_r1": 0,
+        "anli_r2": 0
+    }
+}

4b284b17bc4seed2/evaluation/rankeval/4b284b17bc4seed2_3_lm-eval_global_step80108_2023-02-26-09-23-54_3shots_backup.json

@@ -0,0 +1,16 @@
+{
+    "results": {
+        "anli_r1": {
+            "acc": 0.329,
+            "acc_stderr": 0.014865395385928366
+        },
+        "anli_r2": {
+            "acc": 0.327,
+            "acc_stderr": 0.01484221315341125
+        }
+    },
+    "versions": {
+        "anli_r1": 0,
+        "anli_r2": 0
+    }
+}

4b284b17bc4seed2/evaluation/rankeval/4b284b17bc4seed2_4.json

@@ -0,0 +1,11 @@
+{
+    "results": {
+        "anli_r1": {
+            "acc": 0.332,
+            "acc_stderr": 0.014899597242811487
+        }
+    },
+    "versions": {
+        "anli_r1": 0
+    }
+}

4b284b17bc4seed2/evaluation/rankeval/4b284b17bc4seed2_4_lm-eval_global_step80108_2023-02-26-09-23-54_4shots_backup.json

@@ -0,0 +1,11 @@
+{
+    "results": {
+        "anli_r1": {
+            "acc": 0.332,
+            "acc_stderr": 0.014899597242811487
+        }
+    },
+    "versions": {
+        "anli_r1": 0
+    }
+}

4b284b17bc4seed2/evaluation/rankeval/4b284b17bc4seed2_5.json

@@ -0,0 +1,11 @@
+{
+    "results": {
+        "anli_r1": {
+            "acc": 0.342,
+            "acc_stderr": 0.01500870618212173
+        }
+    },
+    "versions": {
+        "anli_r1": 0
+    }
+}

4b284b17bc4seed2/evaluation/rankeval/4b284b17bc4seed2_5_lm-eval_global_step80108_2023-02-26-09-23-54_5shots_backup.json

@@ -0,0 +1,11 @@
+{
+    "results": {
+        "anli_r1": {
+            "acc": 0.342,
+            "acc_stderr": 0.01500870618212173
+        }
+    },
+    "versions": {
+        "anli_r1": 0
+    }
+}

4b284b17bc4seed2/global_step80108/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d19f4f6717cfa11f4089167db24b0b29e1bb2817134064bb874840168067714c
+size 199058647

4b284b17bc4seed2/global_step80108/bf16_zero_pp_rank_0_mp_rank_01_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:297b32a72bba9aa49c9537d880b9787a5a35eb47181f7deefe32272e4465cef4
+size 199058647

4b284b17bc4seed2/global_step80108/bf16_zero_pp_rank_100_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ea2e6eedb9bc9bd97ec987fb7464df3040b9d475ee75ef384a04ca6722ab85c7
+size 199058733

4b284b17bc4seed2/global_step80108/bf16_zero_pp_rank_100_mp_rank_01_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:18e485b0b011011d03e037d0907626277584d572ad497e420af521594b14d312
+size 199058733

4b284b17bc4seed2/global_step80108/bf16_zero_pp_rank_101_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fd8c889e122a013bb966eb95ba0f0f8234ff81cf1364b05640560685be06e040
+size 199058797

4b284b17bc4seed2/global_step80108/bf16_zero_pp_rank_101_mp_rank_01_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:01595dc0b077c8e72f618d8fd19c9268d9c3a26f2e6b17a1bd3e3ec765d46b09
+size 199058797

4b284b17bc4seed2/global_step80108/bf16_zero_pp_rank_102_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5a4f5a39d584e098799a59a3372814f01ec9420c4ba42739b075b1c068033dd6
+size 199058733

4b284b17bc4seed2/global_step80108/bf16_zero_pp_rank_102_mp_rank_01_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:79c622f72f8dd7644bc87c234d480488be11ba94b1ba0ef4aa7ad95ff3cbc68d
+size 199058733

4b284b17bc4seed2/global_step80108/bf16_zero_pp_rank_103_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f90660386a82d21ec7cf70295a7c3a408a41512f1f4b4024778660c298b13888
+size 199058733

4b284b17bc4seed2/global_step80108/bf16_zero_pp_rank_103_mp_rank_01_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:02315f4f1c48bbbbd3d3cccd3ad5aea92f6fd004b6d5bd191fe78c442895bb48
+size 199058733

4b284b17bc4seed2/global_step80108/bf16_zero_pp_rank_104_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:48e0ad698781c5c8c63cefc365a36a9adc5156770cf5181b65b0ec09228fd37d
+size 199058797

4b284b17bc4seed2/global_step80108/bf16_zero_pp_rank_104_mp_rank_01_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f2774f3874f8dba8200a6104a5debf22a883f11e0120b4c60666423653e01c6e
+size 199058797

4b284b17bc4seed2/global_step80108/bf16_zero_pp_rank_105_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:28fdd8002c30266f396cb23ed293aab7ec62d6c1470841c93e06dcc8a3db3c24
+size 199058733

4b284b17bc4seed2/global_step80108/bf16_zero_pp_rank_105_mp_rank_01_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ff146a02f2012093b98d6a4d6b43107582c7d22febd5f3bcb987428070b3575f
+size 199058733

4b284b17bc4seed2/global_step80108/bf16_zero_pp_rank_106_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e524180750e5069192c18ed03de65c1423ef06d21ade032cbb2a284fb351d745
+size 199058669

4b284b17bc4seed2/global_step80108/bf16_zero_pp_rank_106_mp_rank_01_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2c84e8469b70d872f85a6dee37e0f435241473def04ecd19c43154d2f3c4b4f0
+size 199058669

4b284b17bc4seed2/global_step80108/bf16_zero_pp_rank_107_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:63eea16c8b46dcac2761d4c1815598f7c3f06ffaeb9414179ad3702a9119fb31
+size 199058797

4b284b17bc4seed2/global_step80108/bf16_zero_pp_rank_107_mp_rank_01_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c34b2fe2d547558d51c69b8a2beb3e108bed088e2c1b34c9c0917272aba4b797
+size 199058797

4b284b17bc4seed2/global_step80108/bf16_zero_pp_rank_108_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fd152095fe9200af67618d01205c2b6750cbc6397927eefc3239e567229a894d
+size 199058733

4b284b17bc4seed2/global_step80108/bf16_zero_pp_rank_108_mp_rank_01_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:33e26143bcc52fd589e27998e0770905dd02c6d0e0688f40e37ce8eb05af2d0b
+size 199058733

4b284b17bc4seed2/global_step80108/bf16_zero_pp_rank_109_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1a5f77213756b50fa3818c545bb0408288e9716392636ef535cecf64bb4a268c
+size 199058733

4b284b17bc4seed2/global_step80108/bf16_zero_pp_rank_109_mp_rank_01_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2712850a994a27a466457b619ca2b2424e662a10ff15fe9a87f4384de75ec13b
+size 199058733

4b284b17bc4seed2/global_step80108/bf16_zero_pp_rank_10_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0f757ac808f925c5090c7eea98f335b981638738c829df8cef1dee8c9bbdb720
+size 199058658

4b284b17bc4seed2/global_step80108/bf16_zero_pp_rank_10_mp_rank_01_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9dea9ecc09ae6dae0b1e7b1373c745e329724bd0b9ff2a07af3b4ff4451f46c1
+size 199058658

4b284b17bc4seed2/global_step80108/bf16_zero_pp_rank_110_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4f213727a59089773f69e4a576a4fb43a8ed8fe2f53d06ba127e533436aed464
+size 199058733

4b284b17bc4seed2/global_step80108/bf16_zero_pp_rank_110_mp_rank_01_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc5d89562938d153423a014a48bdf96cca8a240121e59262de84635df48e1de3
+size 199058733

4b284b17bc4seed2/global_step80108/bf16_zero_pp_rank_111_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:81e2b99c559d05f4acab047632c8aa03a50d4fab2a8546054d856bc676aa3623
+size 199058797

4b284b17bc4seed2/global_step80108/bf16_zero_pp_rank_111_mp_rank_01_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a2dfb6e3f6beaee5bd8a62b45042d3f5b312d0544b917a136da14291a94ee3ae
+size 199058797

4b284b17bc4seed2/global_step80108/bf16_zero_pp_rank_112_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9ec70fb5940e8e482dfad0d1383c93027356292f24b5b11d0299e6f832cc6603
+size 199058605