eval-results/AlekseyKorshuk/pygmalion-6b-vicuna-chatml/results_2023-09-09T13-36-28.958118.json

{
  "config_general": {
    "model_name": "AlekseyKorshuk/pygmalion-6b-vicuna-chatml",
    "model_sha": "ee3ada91a69a194cedfabbfeab98f1499b75cb44",
    "model_size": "11.28 GB",
    "model_dtype": "torch.float16",
    "lighteval_sha": "ff467795ccc45b291b69333c263d5f16abd1fcd9",
    "num_few_shot_default": 0,
    "num_fewshot_seeds": 1,
    "override_batch_size": 1,
    "max_samples": null,
    "job_id": ""
  },
  "results": {
    "harness|drop|3": {
      "em": 0.022336409395973155,
      "em_stderr": 0.001513355825080532,
      "f1": 0.07709626677852371,
      "f1_stderr": 0.0019730522657271923
    },
    "harness|gsm8k|5": {
      "acc": 0.025018953752843062,
      "acc_stderr": 0.004302045046564293
    },
    "harness|winogrande|5": {
      "acc": 0.6306235201262825,
      "acc_stderr": 0.013564470596053525
    },
    "all": {
      "em": 0.022336409395973155,
      "em_stderr": 0.001513355825080532,
      "f1": 0.07709626677852371,
      "f1_stderr": 0.0019730522657271923,
      "acc": 0.3278212369395628,
      "acc_stderr": 0.00893325782130891
    }
  },
  "versions": {
    "harness|drop|3": 1,
    "harness|gsm8k|5": 0,
    "harness|winogrande|5": 0,
    "all": 0
  },
  "config_tasks": {
    "harness|drop": "LM Harness task",
    "harness|gsm8k": "LM Harness task",
    "harness|winogrande": "LM Harness task"
  },
  "summary_tasks": {
    "harness|drop|3": {
      "hashes": {
        "hash_examples": "1d27416e8324e9a3",
        "hash_full_prompts": "a5513ff9a741b385",
        "hash_input_tokens": "f21277d2c2d2e06c",
        "hash_cont_tokens": "d7b7fcd0702ec261"
      },
      "truncated": 382,
      "non-truncated": 9154,
      "padded": 0,
      "non-padded": 9536,
      "effective_few_shots": 3.0,
      "num_truncated_few_shots": 0
    },
    "harness|gsm8k|5": {
      "hashes": {
        "hash_examples": "4c0843a5d99bcfdc",
        "hash_full_prompts": "41d55e83abc0e02d",
        "hash_input_tokens": "3ab9b4c5105492a3",
        "hash_cont_tokens": "f191a21e6082b576"
      },
      "truncated": 0,
      "non-truncated": 1319,
      "padded": 0,
      "non-padded": 1319,
      "effective_few_shots": 5.0,
      "num_truncated_few_shots": 0
    },
    "harness|winogrande|5": {
      "hashes": {
        "hash_examples": "aada0a176fd81218",
        "hash_full_prompts": "c8655cbd12de8409",
        "hash_input_tokens": "84cacac1590bb0a5",
        "hash_cont_tokens": "64ca3ed9b5dacc6e"
      },
      "truncated": 0,
      "non-truncated": 2534,
      "padded": 2426,
      "non-padded": 108,
      "effective_few_shots": 5.0,
      "num_truncated_few_shots": 0
    }
  },
  "summary_general": {
    "hashes": {
      "hash_examples": "9b4d8993161e637d",
      "hash_full_prompts": "08215e527b7e60a5",
      "hash_input_tokens": "26cd3631535039d0",
      "hash_cont_tokens": "de68f7850ee6db3d"
    },
    "total_evaluation_time_secondes": "5156.145171165466",
    "truncated": 382,
    "non-truncated": 13007,
    "padded": 2426,
    "non-padded": 10963,
    "num_truncated_few_shots": 0
  }
}