real2_targeted_logic_math_sft_step040/benchmark_comparison.json

{
  "v6_fixed100": {
    "gsm8k": 24.0,
    "math500": 7.0,
    "arc_challenge": 26.0,
    "folio": 29.0,
    "composite": 21.5
  },
  "candidate_fixed100": {
    "gsm8k": 23.0,
    "math500": 7.0,
    "arc_challenge": 29.0,
    "folio": 35.0,
    "composite": 23.5
  },
  "human_preference_balanced80": {
    "v6": 55.0,
    "candidate": 55.0
  }
}