8b7178b58b/evaluation/generation/slim.8b7178b58b_GEM-wiki_lingua_en_tldr_en_1.json

{
  "results": [
    {
      "task_name": "GEM/wiki_lingua_en",
      "prompt_name": "tldr_en",
      "rouge1_precision": 0.2174735691686053,
      "dataset_path": "GEM/wiki_lingua",
      "dataset_name": "en",
      "subset": null,
      "rouge1_precision_stderr": 0.0024727115624900444
    },
    {
      "task_name": "GEM/wiki_lingua_en",
      "prompt_name": "tldr_en",
      "rouge1_recall": 0.3460178647953164,
      "dataset_path": "GEM/wiki_lingua",
      "dataset_name": "en",
      "subset": null,
      "rouge1_recall_stderr": 0.0029105441665587413
    },
    {
      "task_name": "GEM/wiki_lingua_en",
      "prompt_name": "tldr_en",
      "rouge1_fmeasure": 0.24342094967396746,
      "dataset_path": "GEM/wiki_lingua",
      "dataset_name": "en",
      "subset": null,
      "rouge1_fmeasure_stderr": 0.002061008293601447
    },
    {
      "task_name": "GEM/wiki_lingua_en",
      "prompt_name": "tldr_en",
      "rouge2_precision": 0.05974334784565509,
      "dataset_path": "GEM/wiki_lingua",
      "dataset_name": "en",
      "subset": null,
      "rouge2_precision_stderr": 0.0013520311423251665
    },
    {
      "task_name": "GEM/wiki_lingua_en",
      "prompt_name": "tldr_en",
      "rouge2_recall": 0.09579105477293766,
      "dataset_path": "GEM/wiki_lingua",
      "dataset_name": "en",
      "subset": null,
      "rouge2_recall_stderr": 0.0019333447701627138
    },
    {
      "task_name": "GEM/wiki_lingua_en",
      "prompt_name": "tldr_en",
      "rouge2_fmeasure": 0.06579562483959114,
      "dataset_path": "GEM/wiki_lingua",
      "dataset_name": "en",
      "subset": null,
      "rouge2_fmeasure_stderr": 0.0012229149405673168
    },
    {
      "task_name": "GEM/wiki_lingua_en",
      "prompt_name": "tldr_en",
      "rougeL_precision": 0.15264641094715217,
      "dataset_path": "GEM/wiki_lingua",
      "dataset_name": "en",
      "subset": null,
      "rougeL_precision_stderr": 0.0017989249509303383
    },
    {
      "task_name": "GEM/wiki_lingua_en",
      "prompt_name": "tldr_en",
      "rougeL_recall": 0.24988397509922583,
      "dataset_path": "GEM/wiki_lingua",
      "dataset_name": "en",
      "subset": null,
      "rougeL_recall_stderr": 0.0023954858097757943
    },
    {
      "task_name": "GEM/wiki_lingua_en",
      "prompt_name": "tldr_en",
      "rougeL_fmeasure": 0.1715894965219696,
      "dataset_path": "GEM/wiki_lingua",
      "dataset_name": "en",
      "subset": null,
      "rougeL_fmeasure_stderr": 0.0014647407258985053
    },
    {
      "task_name": "GEM/wiki_lingua_en",
      "prompt_name": "tldr_en",
      "rougeLsum_precision": 0.20485722652733274,
      "dataset_path": "GEM/wiki_lingua",
      "dataset_name": "en",
      "subset": null,
      "rougeLsum_precision_stderr": 0.002345428003149735
    },
    {
      "task_name": "GEM/wiki_lingua_en",
      "prompt_name": "tldr_en",
      "rougeLsum_recall": 0.32704916730419714,
      "dataset_path": "GEM/wiki_lingua",
      "dataset_name": "en",
      "subset": null,
      "rougeLsum_recall_stderr": 0.0027965860033602368
    },
    {
      "task_name": "GEM/wiki_lingua_en",
      "prompt_name": "tldr_en",
      "rougeLsum_fmeasure": 0.2293964977081046,
      "dataset_path": "GEM/wiki_lingua",
      "dataset_name": "en",
      "subset": null,
      "rougeLsum_fmeasure_stderr": 0.0019478558777254497
    },
    {
      "task_name": "GEM/wiki_lingua_en",
      "prompt_name": "tldr_en",
      "bleu": 3.5656470388016652,
      "dataset_path": "GEM/wiki_lingua",
      "dataset_name": "en",
      "subset": null,
      "bleu_stderr": 0.08197311523066556
    }
  ],
  "config": {
    "model": "hf-causal",
    "model_args": "pretrained=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-8b7-178b-c4-repetitions/8b7178b58b/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
  }
}