evaluation_val/mlsum_es/agg.limited=3000.model=p31lossseq_global_step498.task=mlsum_es.templates=layman_summ_es.fewshot=0.batchsize=4.seed=1234.timestamp=2022-09-12T20:37:06.json

{"results": [{"task_name": "mlsum_es", "prompt_name": "layman_summ_es", "bleu": 2.7240547697036885, "fixed_answer_choice_list": null, "dataset_path": "GEM/mlsum", "dataset_name": "es", "subset": "", "prompt_id": "6fc70031-95ab-40fa-9cc7-e6eda42a4833", "prompt_jinja": "My college roommate asked me what this Spanish article meant:\n {{text}}\nSo I recapped it in layman''s terms in Spanish: ||| {{target}}", "prompt_original_task": true, "comment": "", "bleu_stderr": 0.09873871923462486}, {"task_name": "mlsum_es", "prompt_name": "layman_summ_es", "rouge1_precision": 0.20922341031018796, "fixed_answer_choice_list": null, "dataset_path": "GEM/mlsum", "dataset_name": "es", "subset": "", "prompt_id": "6fc70031-95ab-40fa-9cc7-e6eda42a4833", "prompt_jinja": "My college roommate asked me what this Spanish article meant:\n {{text}}\nSo I recapped it in layman''s terms in Spanish: ||| {{target}}", "prompt_original_task": true, "comment": "", "rouge1_precision_stderr": 0.0025375595568407133}, {"task_name": "mlsum_es", "prompt_name": "layman_summ_es", "rouge1_recall": 0.2686473017701524, "fixed_answer_choice_list": null, "dataset_path": "GEM/mlsum", "dataset_name": "es", "subset": "", "prompt_id": "6fc70031-95ab-40fa-9cc7-e6eda42a4833", "prompt_jinja": "My college roommate asked me what this Spanish article meant:\n {{text}}\nSo I recapped it in layman''s terms in Spanish: ||| {{target}}", "prompt_original_task": true, "comment": "", "rouge1_recall_stderr": 0.003243235357501857}, {"task_name": "mlsum_es", "prompt_name": "layman_summ_es", "rouge1_fmeasure": 0.2152514623131688, "fixed_answer_choice_list": null, "dataset_path": "GEM/mlsum", "dataset_name": "es", "subset": "", "prompt_id": "6fc70031-95ab-40fa-9cc7-e6eda42a4833", "prompt_jinja": "My college roommate asked me what this Spanish article meant:\n {{text}}\nSo I recapped it in layman''s terms in Spanish: ||| {{target}}", "prompt_original_task": true, "comment": "", "rouge1_fmeasure_stderr": 0.00235779670443836}, {"task_name": "mlsum_es", "prompt_name": "layman_summ_es", "rouge2_precision": 0.05570540199745807, "fixed_answer_choice_list": null, "dataset_path": "GEM/mlsum", "dataset_name": "es", "subset": "", "prompt_id": "6fc70031-95ab-40fa-9cc7-e6eda42a4833", "prompt_jinja": "My college roommate asked me what this Spanish article meant:\n {{text}}\nSo I recapped it in layman''s terms in Spanish: ||| {{target}}", "prompt_original_task": true, "comment": "", "rouge2_precision_stderr": 0.0017110691355917763}, {"task_name": "mlsum_es", "prompt_name": "layman_summ_es", "rouge2_recall": 0.07438656912468752, "fixed_answer_choice_list": null, "dataset_path": "GEM/mlsum", "dataset_name": "es", "subset": "", "prompt_id": "6fc70031-95ab-40fa-9cc7-e6eda42a4833", "prompt_jinja": "My college roommate asked me what this Spanish article meant:\n {{text}}\nSo I recapped it in layman''s terms in Spanish: ||| {{target}}", "prompt_original_task": true, "comment": "", "rouge2_recall_stderr": 0.0021166092536387365}, {"task_name": "mlsum_es", "prompt_name": "layman_summ_es", "rouge2_fmeasure": 0.05799389954948545, "fixed_answer_choice_list": null, "dataset_path": "GEM/mlsum", "dataset_name": "es", "subset": "", "prompt_id": "6fc70031-95ab-40fa-9cc7-e6eda42a4833", "prompt_jinja": "My college roommate asked me what this Spanish article meant:\n {{text}}\nSo I recapped it in layman''s terms in Spanish: ||| {{target}}", "prompt_original_task": true, "comment": "", "rouge2_fmeasure_stderr": 0.0016274426630508128}, {"task_name": "mlsum_es", "prompt_name": "layman_summ_es", "rougeL_precision": 0.15759931030583763, "fixed_answer_choice_list": null, "dataset_path": "GEM/mlsum", "dataset_name": "es", "subset": "", "prompt_id": "6fc70031-95ab-40fa-9cc7-e6eda42a4833", "prompt_jinja": "My college roommate asked me what this Spanish article meant:\n {{text}}\nSo I recapped it in layman''s terms in Spanish: ||| {{target}}", "prompt_original_task": true, "comment": "", "rougeL_precision_stderr": 0.002106134774825538}, {"task_name": "mlsum_es", "prompt_name": "layman_summ_es", "rougeL_recall": 0.19930722307414317, "fixed_answer_choice_list": null, "dataset_path": "GEM/mlsum", "dataset_name": "es", "subset": "", "prompt_id": "6fc70031-95ab-40fa-9cc7-e6eda42a4833", "prompt_jinja": "My college roommate asked me what this Spanish article meant:\n {{text}}\nSo I recapped it in layman''s terms in Spanish: ||| {{target}}", "prompt_original_task": true, "comment": "", "rougeL_recall_stderr": 0.002532772355449645}, {"task_name": "mlsum_es", "prompt_name": "layman_summ_es", "rougeL_fmeasure": 0.16039478846940786, "fixed_answer_choice_list": null, "dataset_path": "GEM/mlsum", "dataset_name": "es", "subset": "", "prompt_id": "6fc70031-95ab-40fa-9cc7-e6eda42a4833", "prompt_jinja": "My college roommate asked me what this Spanish article meant:\n {{text}}\nSo I recapped it in layman''s terms in Spanish: ||| {{target}}", "prompt_original_task": true, "comment": "", "rougeL_fmeasure_stderr": 0.001896226389379961}, {"task_name": "mlsum_es", "prompt_name": "layman_summ_es", "rougeLsum_precision": 0.16395293631004723, "fixed_answer_choice_list": null, "dataset_path": "GEM/mlsum", "dataset_name": "es", "subset": "", "prompt_id": "6fc70031-95ab-40fa-9cc7-e6eda42a4833", "prompt_jinja": "My college roommate asked me what this Spanish article meant:\n {{text}}\nSo I recapped it in layman''s terms in Spanish: ||| {{target}}", "prompt_original_task": true, "comment": "", "rougeLsum_precision_stderr": 0.002154082917851902}, {"task_name": "mlsum_es", "prompt_name": "layman_summ_es", "rougeLsum_recall": 0.20952028063185008, "fixed_answer_choice_list": null, "dataset_path": "GEM/mlsum", "dataset_name": "es", "subset": "", "prompt_id": "6fc70031-95ab-40fa-9cc7-e6eda42a4833", "prompt_jinja": "My college roommate asked me what this Spanish article meant:\n {{text}}\nSo I recapped it in layman''s terms in Spanish: ||| {{target}}", "prompt_original_task": true, "comment": "", "rougeLsum_recall_stderr": 0.0026942495945950105}, {"task_name": "mlsum_es", "prompt_name": "layman_summ_es", "rougeLsum_fmeasure": 0.16743568643555828, "fixed_answer_choice_list": null, "dataset_path": "GEM/mlsum", "dataset_name": "es", "subset": "", "prompt_id": "6fc70031-95ab-40fa-9cc7-e6eda42a4833", "prompt_jinja": "My college roommate asked me what this Spanish article meant:\n {{text}}\nSo I recapped it in layman''s terms in Spanish: ||| {{target}}", "prompt_original_task": true, "comment": "", "rougeLsum_fmeasure_stderr": 0.0019537523119573046}], "config": {"model": "hf-causal", "model_args": "pretrained=/gpfsscratch/rech/six/commun/commun/experiments/muennighoff/bloomckpt/176bt0/p31lossseq_global_step498,use_accelerate=True,tokenizer=/gpfsscratch/rech/six/commun/commun/experiments/muennighoff/bloomckpt/176bt0/p31lossseq_global_step498,dtype=bfloat16", "num_fewshot": 0, "batch_size": 4, "device": "cuda", "use_cache": false, "limit": 3000, "bootstrap_iters": 10, "seed": 1234}}