[
  {
    "ckpt_id": "ckpt-q05-lr1e4-r16-wikiqa1-s42",
    "benchmark": "kobest_boolq",
    "eval_set_id": "kobest_boolq-test",
    "score": 75.1
  },
  {
    "ckpt_id": "ckpt-q05-lr1e4-r16-wikiqa1-s42",
    "benchmark": "gsm8k",
    "eval_set_id": "gsm8k-test",
    "score": 40.7
  },
  {
    "ckpt_id": "ckpt-q05-lr1e4-r16-wikiqa2-s1",
    "benchmark": "kobest_boolq",
    "eval_set_id": "kobest_boolq-test",
    "score": 67.2
  },
  {
    "ckpt_id": "ckpt-q05-lr1e4-r16-wikiqa2-s1",
    "benchmark": "gsm8k",
    "eval_set_id": "gsm8k-test",
    "score": 40.9
  },
  {
    "ckpt_id": "ckpt-q05-lr1e4-r16-wikiqa2-s42",
    "benchmark": "kobest_boolq",
    "eval_set_id": "kobest_boolq-test",
    "score": 68
  },
  {
    "ckpt_id": "ckpt-q05-lr1e4-r16-wikiqa2-s42",
    "benchmark": "gsm8k",
    "eval_set_id": "gsm8k-test",
    "score": 40.3
  },
  {
    "ckpt_id": "ckpt-q05-lr3e4-r64-wikiqa2-s1",
    "benchmark": "kobest_boolq",
    "eval_set_id": "kobest_boolq-test",
    "score": 64.1
  },
  {
    "ckpt_id": "ckpt-q05-lr3e4-r64-wikiqa2-s1",
    "benchmark": "gsm8k",
    "eval_set_id": "gsm8k-test",
    "score": 40.6
  },
  {
    "ckpt_id": "ckpt-q05-lr3e4-r64-wikiqa2-s7",
    "benchmark": "kobest_boolq",
    "eval_set_id": "kobest_boolq-test",
    "score": 69.8
  },
  {
    "ckpt_id": "ckpt-q05-lr3e4-r64-wikiqa2-s7",
    "benchmark": "gsm8k",
    "eval_set_id": "gsm8k-test",
    "score": 40.1
  },
  {
    "ckpt_id": "ckpt-q05-lr3e4-r64-wikiqa2-s42",
    "benchmark": "kobest_boolq",
    "eval_set_id": "kobest_boolq-test",
    "score": 72.3
  },
  {
    "ckpt_id": "ckpt-q05-lr3e4-r64-wikiqa2-s42",
    "benchmark": "gsm8k",
    "eval_set_id": "gsm8k-test",
    "score": 39.7
  },
  {
    "ckpt_id": "ckpt-q05-lr1e4-r16-wikiqa2-ep3-s42",
    "benchmark": "kobest_boolq",
    "eval_set_id": "kobest_boolq-test",
    "score": 71.2
  },
  {
    "ckpt_id": "ckpt-q05-lr1e4-r16-wikiqa2-ep3-s42",
    "benchmark": "gsm8k",
    "eval_set_id": "gsm8k-test",
    "score": 37.5
  },
  {
    "ckpt_id": "ckpt-q05-lr1e4-r16-wikiqa2b2-s42",
    "benchmark": "kobest_boolq",
    "eval_set_id": "kobest_boolq-test",
    "score": 69.4
  },
  {
    "ckpt_id": "ckpt-q05-lr1e4-r16-wikiqa2b2-s42",
    "benchmark": "gsm8k",
    "eval_set_id": "gsm8k-test",
    "score": 40.5
  },
  {
    "ckpt_id": "ckpt-q15-lr1e4-r16-wikiqa2-s42",
    "benchmark": "kobest_boolq",
    "eval_set_id": "kobest_boolq-test",
    "score": 78.9
  },
  {
    "ckpt_id": "ckpt-q15-lr1e4-r16-wikiqa2-s42",
    "benchmark": "gsm8k",
    "eval_set_id": "gsm8k-test",
    "score": 62.2
  },
  {
    "ckpt_id": "ckpt-q15-lr1e4-r16-wikiqa2-s7",
    "benchmark": "kobest_boolq",
    "eval_set_id": "kobest_boolq-test",
    "score": 78.1
  },
  {
    "ckpt_id": "ckpt-q15-lr1e4-r16-wikiqa2-s7",
    "benchmark": "gsm8k",
    "eval_set_id": "gsm8k-test",
    "score": 62.1
  }
]