evaluation_val/mlsum_es/slim.limited=3000.model=p31lossseq_global_step498.task=mlsum_es.templates=palm_prompt.fewshot=0.batchsize=4.seed=1234.timestamp=2022-09-12T20:37:19.json

{
  "results": [
    {
      "task_name": "mlsum_es",
      "prompt_name": "palm_prompt",
      "bleu": 3.916653973278921,
      "dataset_path": "GEM/mlsum",
      "dataset_name": "es",
      "subset": "",
      "bleu_stderr": 0.13320174843239593
    },
    {
      "task_name": "mlsum_es",
      "prompt_name": "palm_prompt",
      "rouge1_precision": 0.238006133260283,
      "dataset_path": "GEM/mlsum",
      "dataset_name": "es",
      "subset": "",
      "rouge1_precision_stderr": 0.002592201534618347
    },
    {
      "task_name": "mlsum_es",
      "prompt_name": "palm_prompt",
      "rouge1_recall": 0.28440217742758356,
      "dataset_path": "GEM/mlsum",
      "dataset_name": "es",
      "subset": "",
      "rouge1_recall_stderr": 0.0032686854143649984
    },
    {
      "task_name": "mlsum_es",
      "prompt_name": "palm_prompt",
      "rouge1_fmeasure": 0.23777639529079925,
      "dataset_path": "GEM/mlsum",
      "dataset_name": "es",
      "subset": "",
      "rouge1_fmeasure_stderr": 0.002353644348181871
    },
    {
      "task_name": "mlsum_es",
      "prompt_name": "palm_prompt",
      "rouge2_precision": 0.07002703238460953,
      "dataset_path": "GEM/mlsum",
      "dataset_name": "es",
      "subset": "",
      "rouge2_precision_stderr": 0.0019467819130961877
    },
    {
      "task_name": "mlsum_es",
      "prompt_name": "palm_prompt",
      "rouge2_recall": 0.08785375976968864,
      "dataset_path": "GEM/mlsum",
      "dataset_name": "es",
      "subset": "",
      "rouge2_recall_stderr": 0.002327977791822037
    },
    {
      "task_name": "mlsum_es",
      "prompt_name": "palm_prompt",
      "rouge2_fmeasure": 0.07113316160241126,
      "dataset_path": "GEM/mlsum",
      "dataset_name": "es",
      "subset": "",
      "rouge2_fmeasure_stderr": 0.0018688817237639986
    },
    {
      "task_name": "mlsum_es",
      "prompt_name": "palm_prompt",
      "rougeL_precision": 0.18083602836591542,
      "dataset_path": "GEM/mlsum",
      "dataset_name": "es",
      "subset": "",
      "rougeL_precision_stderr": 0.002222442145519105
    },
    {
      "task_name": "mlsum_es",
      "prompt_name": "palm_prompt",
      "rougeL_recall": 0.21460635266053055,
      "dataset_path": "GEM/mlsum",
      "dataset_name": "es",
      "subset": "",
      "rougeL_recall_stderr": 0.0026818176382247703
    },
    {
      "task_name": "mlsum_es",
      "prompt_name": "palm_prompt",
      "rougeL_fmeasure": 0.17988536856242504,
      "dataset_path": "GEM/mlsum",
      "dataset_name": "es",
      "subset": "",
      "rougeL_fmeasure_stderr": 0.0020214064930605237
    },
    {
      "task_name": "mlsum_es",
      "prompt_name": "palm_prompt",
      "rougeLsum_precision": 0.18813961119405384,
      "dataset_path": "GEM/mlsum",
      "dataset_name": "es",
      "subset": "",
      "rougeLsum_precision_stderr": 0.00226738528856828
    },
    {
      "task_name": "mlsum_es",
      "prompt_name": "palm_prompt",
      "rougeLsum_recall": 0.2256886585459008,
      "dataset_path": "GEM/mlsum",
      "dataset_name": "es",
      "subset": "",
      "rougeLsum_recall_stderr": 0.0028543730103714284
    },
    {
      "task_name": "mlsum_es",
      "prompt_name": "palm_prompt",
      "rougeLsum_fmeasure": 0.18777680808178215,
      "dataset_path": "GEM/mlsum",
      "dataset_name": "es",
      "subset": "",
      "rougeLsum_fmeasure_stderr": 0.0020733063902455624
    }
  ],
  "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
  }
}