R596 — SoftCtx × HiRank × MidBeta MegaExtra (offline DPO)

How this checkpoint was trained

• Base / parent: unconst/Affine-5czsc2fc98-r252-merged@b42d6245d77fe30885ea8a90387771e1bc465e0f (our crowned r252, reign 33)
• Method: offline DPO on Reason-ranked pairs (not SFT / not online GRPO)
• What was optimized: preference for higher teacher-side Reason on mined pairs
• Data: SoftCtx × HiRank pair set (soft context length band, high-rank filter); MidBeta β=0.1. See experiment plan.md / dpo_duel_reason.jsonl under mining/experiments/r596-r252-offline-dpo-hialpha-hirank-midbeta-softctx-megaextrasteps.
• Key hyperparameters:
  • LoRA r=64, α=128
  • β=0.1 (MidBeta)
  • lr=5e-6
  • max_len=12288 (SoftCtx)
  • max_steps target 3600 MegaExtra; TRAIN_DONE@259 (adapter kept / merged)
• Hardware: Lium mine-r226-marsplan-fullft-1 (brave) GPUs 6,7 for train; merge + n80 on mine-r252-vera-t4-nonking-grpo-1 GPUs 4,5 → /tmp/r596_merged
• Local n80 vs live king reign34 (cryptoDev23/Affine-5Dku3dYp9j-hk8161@55b7ffe0…):
  • margin +0.006196, SE 0.002357, z=2.63, n=75
  • bar max(2·SE, δ=0.002) = 0.004713 (~1.31×)
  • thought median 199 (≥80), B pass 0.368 (≥0.30)
  • decision: Stage-5 licensed (r596_decision_reign34.json)
• Prior n80 vs r252: margin +0.008490 (~1.19× bar), thought/B clear
• Experiment path: mining/experiments/r596-r252-offline-dpo-hialpha-hirank-midbeta-softctx-megaextrasteps

Intended use

License