{
  "results": {
    "kobest_boolq": {
      "acc": 0.5028490028490028,
      "acc_stderr": 0.013348550797680814,
      "macro_f1": 0.3371320037986705,
      "macro_f1_stderr": 0.006306119104194062
    },
    "kobest_copa": {
      "acc": 0.766,
      "acc_stderr": 0.013394902889660007,
      "macro_f1": 0.7652638674884438,
      "macro_f1_stderr": 0.01339434219317699
    },
    "kobest_hellaswag": {
      "acc": 0.444,
      "acc_stderr": 0.02224224437573102,
      "acc_norm": 0.552,
      "acc_norm_stderr": 0.02226169729227013,
      "macro_f1": 0.4425279610634302,
      "macro_f1_stderr": 0.02226671964725493
    },
    "kobest_sentineg": {
      "acc": 0.5037783375314862,
      "acc_stderr": 0.025125227983562776,
      "macro_f1": 0.33500837520938026,
      "macro_f1_stderr": 0.011139659523368081
    }
  },
  "versions": {
    "kobest_boolq": 0,
    "kobest_copa": 0,
    "kobest_hellaswag": 0,
    "kobest_sentineg": 0
  },
  "config": {
    "model": "gpt2",
    "model_args": "pretrained=nlpai-lab/kullm-polyglot-5.8b-v2",
    "num_fewshot": 0,
    "batch_size": "2",
    "device": "cuda:0",
    "no_cache": true,
    "limit": null,
    "bootstrap_iters": 100000,
    "description_dict": {}
  }
}