Views
No views yet
<thought> block to implement 3-persona internal debate:| Weight | Function | Signal |
|---|---|---|
| 0.40 | consensus_correctness | +1.5 correct consensus, -1.0 wrong, None malformed |
| 0.25 | logic_leap_punishment | -2.0 if FC critique ignored, +0.5 if addressed |
| 0.20 | rmad_format_reward | Graduated 0→1.0 for structural compliance |
| 0.10 | debate_depth_reward | Multi-round substantive debate quality |
| 0.05 | soft_overlong | Length penalty (DAPO-style) |
accelerate launch rmad_grpo_train.py