bench_results.json

{
  "method": "full",
  "seed": 0,
  "train_seconds": 726.3,
  "peak_mem_gb": 23.04,
  "humaneval": 0.524390243902439,
  "humaneval_plus": 0.4817073170731707,
  "mbpp": 0.5264550264550265,
  "mbpp_plus": 0.4497354497354497
}