v1/evaluation/slim.limited=3000.model=lm1-2b8-55b-c4-perplexity.task=e2e_nlg_cleaned.templates=generate_text_restaurant.fewshot=1.batchsize=16.seed=1234.timestamp=2023-01-17T12:14:07.json

{
  "results": [
    {
      "task_name": "e2e_nlg_cleaned",
      "prompt_name": "generate_text_restaurant",
      "bleu": 4.60857341288916,
      "dataset_path": "e2e_nlg_cleaned",
      "dataset_name": null,
      "subset": null,
      "bleu_stderr": 0.07091478097604244
    },
    {
      "task_name": "e2e_nlg_cleaned",
      "prompt_name": "generate_text_restaurant",
      "rouge1_precision": 0.22003605992748757,
      "dataset_path": "e2e_nlg_cleaned",
      "dataset_name": null,
      "subset": null,
      "rouge1_precision_stderr": 0.0016634925664108194
    },
    {
      "task_name": "e2e_nlg_cleaned",
      "prompt_name": "generate_text_restaurant",
      "rouge1_recall": 0.4349109426362562,
      "dataset_path": "e2e_nlg_cleaned",
      "dataset_name": null,
      "subset": null,
      "rouge1_recall_stderr": 0.0026795975680511614
    },
    {
      "task_name": "e2e_nlg_cleaned",
      "prompt_name": "generate_text_restaurant",
      "rouge1_fmeasure": 0.28307413792452757,
      "dataset_path": "e2e_nlg_cleaned",
      "dataset_name": null,
      "subset": null,
      "rouge1_fmeasure_stderr": 0.0017345666507319096
    },
    {
      "task_name": "e2e_nlg_cleaned",
      "prompt_name": "generate_text_restaurant",
      "rouge2_precision": 0.08133706936294005,
      "dataset_path": "e2e_nlg_cleaned",
      "dataset_name": null,
      "subset": null,
      "rouge2_precision_stderr": 0.001011706981693919
    },
    {
      "task_name": "e2e_nlg_cleaned",
      "prompt_name": "generate_text_restaurant",
      "rouge2_recall": 0.16609374412846958,
      "dataset_path": "e2e_nlg_cleaned",
      "dataset_name": null,
      "subset": null,
      "rouge2_recall_stderr": 0.001955786555953094
    },
    {
      "task_name": "e2e_nlg_cleaned",
      "prompt_name": "generate_text_restaurant",
      "rouge2_fmeasure": 0.10538118293432128,
      "dataset_path": "e2e_nlg_cleaned",
      "dataset_name": null,
      "subset": null,
      "rouge2_fmeasure_stderr": 0.0011922460727243065
    },
    {
      "task_name": "e2e_nlg_cleaned",
      "prompt_name": "generate_text_restaurant",
      "rougeL_precision": 0.17905562119769688,
      "dataset_path": "e2e_nlg_cleaned",
      "dataset_name": null,
      "subset": null,
      "rougeL_precision_stderr": 0.0011815805464116809
    },
    {
      "task_name": "e2e_nlg_cleaned",
      "prompt_name": "generate_text_restaurant",
      "rougeL_recall": 0.36101365402052843,
      "dataset_path": "e2e_nlg_cleaned",
      "dataset_name": null,
      "subset": null,
      "rougeL_recall_stderr": 0.002253209383791391
    },
    {
      "task_name": "e2e_nlg_cleaned",
      "prompt_name": "generate_text_restaurant",
      "rougeL_fmeasure": 0.2320488807632003,
      "dataset_path": "e2e_nlg_cleaned",
      "dataset_name": null,
      "subset": null,
      "rougeL_fmeasure_stderr": 0.001266631291357149
    },
    {
      "task_name": "e2e_nlg_cleaned",
      "prompt_name": "generate_text_restaurant",
      "rougeLsum_precision": 0.17784834223848153,
      "dataset_path": "e2e_nlg_cleaned",
      "dataset_name": null,
      "subset": null,
      "rougeLsum_precision_stderr": 0.001488841924808455
    },
    {
      "task_name": "e2e_nlg_cleaned",
      "prompt_name": "generate_text_restaurant",
      "rougeLsum_recall": 0.3521512097724569,
      "dataset_path": "e2e_nlg_cleaned",
      "dataset_name": null,
      "subset": null,
      "rougeLsum_recall_stderr": 0.0025235785084938953
    },
    {
      "task_name": "e2e_nlg_cleaned",
      "prompt_name": "generate_text_restaurant",
      "rougeLsum_fmeasure": 0.22885806406725848,
      "dataset_path": "e2e_nlg_cleaned",
      "dataset_name": null,
      "subset": null,
      "rougeLsum_fmeasure_stderr": 0.0016097640905790775
    }
  ],
  "config": {
    "model": "hf-causal",
    "model_args": "pretrained=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-2b8-55b-c4-perplexity/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
  }
}