eval_results/HuggingFaceH4/qwen-1.5-1.8b-dpo/v0.1/arc/results_2024-03-22T22-03-06.534187.json

{
  "config_general": {
    "lighteval_sha": "?",
    "num_fewshot_seeds": 1,
    "override_batch_size": 1,
    "max_samples": null,
    "job_id": "",
    "start_time": 46958.496830519,
    "end_time": 47032.577119774,
    "total_evaluation_time_secondes": "74.08028925500548",
    "model_name": "HuggingFaceH4/qwen-1.5-1.8b-dpo",
    "model_sha": "fb7ed2c50e9b54d192d279f74fc953226848740b",
    "model_dtype": "torch.bfloat16",
    "model_size": "3.79 GB",
    "config": null
  },
  "results": {
    "leaderboard|arc:challenge|25": {
      "acc": 0.36945392491467577,
      "acc_stderr": 0.014104578366491897,
      "acc_norm": 0.39505119453924914,
      "acc_norm_stderr": 0.014285898292938169
    }
  },
  "versions": {
    "leaderboard|arc:challenge|25": 0
  },
  "config_tasks": {
    "leaderboard|arc:challenge": {
      "name": "arc:challenge",
      "prompt_function": "arc",
      "hf_repo": "ai2_arc",
      "hf_subset": "ARC-Challenge",
      "metric": [
        "loglikelihood_acc",
        "loglikelihood_acc_norm_nospace"
      ],
      "hf_avail_splits": [
        "train",
        "test"
      ],
      "evaluation_splits": [
        "test"
      ],
      "few_shots_split": null,
      "few_shots_select": "random_sampling_from_train",
      "generation_size": 1,
      "stop_sequence": [
        "\n"
      ],
      "output_regex": null,
      "frozen": false,
      "suite": [
        "leaderboard",
        "arc"
      ],
      "original_num_docs": 1172,
      "effective_num_docs": 1172,
      "trust_dataset": true
    }
  },
  "summary_tasks": {
    "leaderboard|arc:challenge|25": {
      "hashes": {
        "hash_examples": "17b0cae357c0259e",
        "hash_full_prompts": "66498cebc864f4ca",
        "hash_input_tokens": "00488c5938fefb2d",
        "hash_cont_tokens": "da3689055cb5fa28"
      },
      "truncated": 0,
      "non_truncated": 1172,
      "padded": 4651,
      "non_padded": 36,
      "effective_few_shots": 25.0,
      "num_truncated_few_shots": 0
    }
  },
  "summary_general": {
    "hashes": {
      "hash_examples": "aaa6929c6d3771fb",
      "hash_full_prompts": "848db321e3b0d4cd",
      "hash_input_tokens": "afc857c7cc852ee9",
      "hash_cont_tokens": "da91173bce47858f"
    },
    "truncated": 0,
    "non_truncated": 1172,
    "padded": 4651,
    "non_padded": 36,
    "num_truncated_few_shots": 0
  }
}