4b284b12bc4/evaluation/generation/slim.4b284b12bc4_GEM-wiki_lingua_en_tldr_en_1.json

{
  "results": [
    {
      "task_name": "GEM/wiki_lingua_en",
      "prompt_name": "tldr_en",
      "rouge1_precision": 0.13584035826185337,
      "dataset_path": "GEM/wiki_lingua",
      "dataset_name": "en",
      "subset": null,
      "rouge1_precision_stderr": 0.001883003830967405
    },
    {
      "task_name": "GEM/wiki_lingua_en",
      "prompt_name": "tldr_en",
      "rouge1_recall": 0.2300655738827389,
      "dataset_path": "GEM/wiki_lingua",
      "dataset_name": "en",
      "subset": null,
      "rouge1_recall_stderr": 0.002771815067524841
    },
    {
      "task_name": "GEM/wiki_lingua_en",
      "prompt_name": "tldr_en",
      "rouge1_fmeasure": 0.15836293018887368,
      "dataset_path": "GEM/wiki_lingua",
      "dataset_name": "en",
      "subset": null,
      "rouge1_fmeasure_stderr": 0.001891640433790272
    },
    {
      "task_name": "GEM/wiki_lingua_en",
      "prompt_name": "tldr_en",
      "rouge2_precision": 0.02400583408187123,
      "dataset_path": "GEM/wiki_lingua",
      "dataset_name": "en",
      "subset": null,
      "rouge2_precision_stderr": 0.0007285343955308211
    },
    {
      "task_name": "GEM/wiki_lingua_en",
      "prompt_name": "tldr_en",
      "rouge2_recall": 0.04289680599794199,
      "dataset_path": "GEM/wiki_lingua",
      "dataset_name": "en",
      "subset": null,
      "rouge2_recall_stderr": 0.001353605133009988
    },
    {
      "task_name": "GEM/wiki_lingua_en",
      "prompt_name": "tldr_en",
      "rouge2_fmeasure": 0.028190707681194575,
      "dataset_path": "GEM/wiki_lingua",
      "dataset_name": "en",
      "subset": null,
      "rouge2_fmeasure_stderr": 0.0008118715649523407
    },
    {
      "task_name": "GEM/wiki_lingua_en",
      "prompt_name": "tldr_en",
      "rougeL_precision": 0.10298984208878467,
      "dataset_path": "GEM/wiki_lingua",
      "dataset_name": "en",
      "subset": null,
      "rougeL_precision_stderr": 0.0012746961770117027
    },
    {
      "task_name": "GEM/wiki_lingua_en",
      "prompt_name": "tldr_en",
      "rougeL_recall": 0.18009351587124983,
      "dataset_path": "GEM/wiki_lingua",
      "dataset_name": "en",
      "subset": null,
      "rougeL_recall_stderr": 0.002104050870387258
    },
    {
      "task_name": "GEM/wiki_lingua_en",
      "prompt_name": "tldr_en",
      "rougeL_fmeasure": 0.12130392597137107,
      "dataset_path": "GEM/wiki_lingua",
      "dataset_name": "en",
      "subset": null,
      "rougeL_fmeasure_stderr": 0.0012906736654645788
    },
    {
      "task_name": "GEM/wiki_lingua_en",
      "prompt_name": "tldr_en",
      "rougeLsum_precision": 0.12668865175777289,
      "dataset_path": "GEM/wiki_lingua",
      "dataset_name": "en",
      "subset": null,
      "rougeLsum_precision_stderr": 0.0017429813042225584
    },
    {
      "task_name": "GEM/wiki_lingua_en",
      "prompt_name": "tldr_en",
      "rougeLsum_recall": 0.21558088869876474,
      "dataset_path": "GEM/wiki_lingua",
      "dataset_name": "en",
      "subset": null,
      "rougeLsum_recall_stderr": 0.0026000978058887433
    },
    {
      "task_name": "GEM/wiki_lingua_en",
      "prompt_name": "tldr_en",
      "rougeLsum_fmeasure": 0.14788907195330203,
      "dataset_path": "GEM/wiki_lingua",
      "dataset_name": "en",
      "subset": null,
      "rougeLsum_fmeasure_stderr": 0.0017520956258023405
    },
    {
      "task_name": "GEM/wiki_lingua_en",
      "prompt_name": "tldr_en",
      "bleu": 1.346869321685321,
      "dataset_path": "GEM/wiki_lingua",
      "dataset_name": "en",
      "subset": null,
      "bleu_stderr": 0.06607443264134674
    }
  ],
  "config": {
    "model": "hf-causal",
    "model_args": "pretrained=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4-repetitions/4b284b12bc4/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
  }
}