eval_results/HuggingFaceH4/Qwen1.5-1.8B-Chat-dpo/v0.1/winogrande/results_2024-03-22T14-04-59.543259.json

{
  "config_general": {
    "lighteval_sha": "?",
    "num_fewshot_seeds": 1,
    "override_batch_size": 1,
    "max_samples": null,
    "job_id": "",
    "start_time": 862626.790740503,
    "end_time": 862679.4968283,
    "total_evaluation_time_secondes": "52.70608779694885",
    "model_name": "HuggingFaceH4/Qwen1.5-1.8B-Chat-dpo",
    "model_sha": "b265371df32e84a5e9938b5a33b1adfd9db4ebd4",
    "model_dtype": "torch.bfloat16",
    "model_size": "3.8 GB",
    "config": null
  },
  "results": {
    "leaderboard|winogrande|5": {
      "acc": 0.5540647198105761,
      "acc_stderr": 0.013970093482330704
    }
  },
  "versions": {
    "leaderboard|winogrande|5": 0
  },
  "config_tasks": {
    "leaderboard|winogrande": {
      "name": "winogrande",
      "prompt_function": "winogrande",
      "hf_repo": "winogrande",
      "hf_subset": "winogrande_xl",
      "metric": [
        "loglikelihood_acc"
      ],
      "hf_avail_splits": [
        "train",
        "test",
        "validation"
      ],
      "evaluation_splits": [
        "validation"
      ],
      "few_shots_split": null,
      "few_shots_select": "random_sampling",
      "generation_size": -1,
      "stop_sequence": [
        "\n"
      ],
      "output_regex": null,
      "frozen": false,
      "suite": [
        "leaderboard"
      ],
      "original_num_docs": 1267,
      "effective_num_docs": 1267,
      "trust_dataset": true
    }
  },
  "summary_tasks": {
    "leaderboard|winogrande|5": {
      "hashes": {
        "hash_examples": "087d5d1a1afd4c7b",
        "hash_full_prompts": "a62b027af32f1fe4",
        "hash_input_tokens": "2d5322eff4ad78c0",
        "hash_cont_tokens": "af83e7e371fcf4d4"
      },
      "truncated": 0,
      "non_truncated": 1267,
      "padded": 2356,
      "non_padded": 178,
      "effective_few_shots": 5.0,
      "num_truncated_few_shots": 0
    }
  },
  "summary_general": {
    "hashes": {
      "hash_examples": "b9a49975cc41fab7",
      "hash_full_prompts": "1d1bae16e80ffe07",
      "hash_input_tokens": "4228ab07fdbbb16c",
      "hash_cont_tokens": "cf29ee9ce86b26c4"
    },
    "truncated": 0,
    "non_truncated": 1267,
    "padded": 2356,
    "non_padded": 178,
    "num_truncated_few_shots": 0
  }
}