{
  "results": {
    "kobest_boolq": {
      "acc": 0.5256410256410257,
      "acc_stderr": 0.013331203323083502,
      "macro_f1": 0.3992444490131579,
      "macro_f1_stderr": 0.010971221905948464
    },
    "kobest_copa": {
      "acc": 0.785,
      "acc_stderr": 0.012997843819031832,
      "macro_f1": 0.7847052615029975,
      "macro_f1_stderr": 0.012967330755596378
    },
    "kobest_hellaswag": {
      "acc": 0.466,
      "acc_stderr": 0.02233126442325838,
      "acc_norm": 0.592,
      "acc_norm_stderr": 0.022000910893877193,
      "macro_f1": 0.46440112486053614,
      "macro_f1_stderr": 0.022343780241897912
    },
    "kobest_sentineg": {
      "acc": 0.5062972292191436,
      "acc_stderr": 0.02512395255890725,
      "macro_f1": 0.34501683501683506,
      "macro_f1_stderr": 0.01352473866949455
    }
  },
  "versions": {
    "kobest_boolq": 0,
    "kobest_copa": 0,
    "kobest_hellaswag": 0,
    "kobest_sentineg": 0
  },
  "config": {
    "model": "gpt2",
    "model_args": "pretrained=jojo0217/ChatSKKU5.8B",
    "num_fewshot": 0,
    "batch_size": "4",
    "device": "cuda:0",
    "no_cache": true,
    "limit": null,
    "bootstrap_iters": 100000,
    "description_dict": {}
  }
}