meta-llama/Llama-2-7b-hf/results_2023-09-09T12-32-30.613622.json

{
  "config_general": {
    "model_name": "meta-llama/Llama-2-7b-hf",
    "model_sha": "6fdf2e60f86ff2481f2241aaee459f85b5b0bbb9",
    "model_size": "3.57 GB",
    "model_dtype": "4bit",
    "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": 0.0010486577181208054,
      "em_stderr": 0.00033145814652191404,
      "f1": 0.05131291946308739,
      "f1_stderr": 0.0012542058656851648
    },
    "harness|gsm8k|5": {
      "acc": 0.053828658074298714,
      "acc_stderr": 0.006216328640238123
    },
    "harness|winogrande|5": {
      "acc": 0.7458563535911602,
      "acc_stderr": 0.012236307219708262
    },
    "all": {
      "em": 0.0010486577181208054,
      "em_stderr": 0.00033145814652191404,
      "f1": 0.05131291946308739,
      "f1_stderr": 0.0012542058656851648,
      "acc": 0.39984250583272946,
      "acc_stderr": 0.009226317929973193
    }
  },
  "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": "42076f0efbb50aa6",
        "hash_cont_tokens": "586467d69620d89a"
      },
      "truncated": 3,
      "non-truncated": 9533,
      "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": "bda342e47b5099b2",
        "hash_cont_tokens": "d10e414c90dc2a07"
      },
      "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": "c0bedf98cb040854",
        "hash_cont_tokens": "f08975ad6f2d5864"
      },
      "truncated": 0,
      "non-truncated": 2534,
      "padded": 2432,
      "non-padded": 102,
      "effective_few_shots": 5.0,
      "num_truncated_few_shots": 0
    }
  },
  "summary_general": {
    "hashes": {
      "hash_examples": "9b4d8993161e637d",
      "hash_full_prompts": "08215e527b7e60a5",
      "hash_input_tokens": "a12f3e3c934bd78b",
      "hash_cont_tokens": "a4df4eaec2b09094"
    },
    "total_evaluation_time_secondes": "8584.129543542862",
    "truncated": 3,
    "non-truncated": 13386,
    "padded": 2432,
    "non-padded": 10957,
    "num_truncated_few_shots": 0
  }
}