results.json

{
  "standard": {
    "HumanEval": 0.61,
    "MBPP": 3.5
  },
  "reasoning": {
    "HumanEval": 0.0,
    "MBPP": 0.5
  },
  "meta": {
    "params_M": 212.4,
    "context": 2048,
    "he_n": 164,
    "mbpp_n": 200,
    "refine_iters": 2,
    "minutes": 11.4
  }
}