data/eval_grpo.json

{
  "name": "model[checkpoints/grpo_final]",
  "episodes": 1000,
  "logical_correction_rate": 0.964,
  "pymatching_beat_rate": 0.0,
  "format_compliance_rate": 1.0,
  "format_partial_rate": 0.0,
  "syndrome_consistency_rate": 0.734,
  "mean_syndrome_consistency": 0.867,
  "mean_hamming_overlap": 0.8405,
  "mean_total_reward": 0.8209249999999999,
  "exact_match_pymatching": 0.734,
  "mean_output_length": 7.0,
  "level": "L2_target"
}