Affine-5czsc2fc98-r576-r252-odpo-hirank-shortctx-ultraextra-merged

Training story

• Base / parent king: unconst/Affine-5czsc2fc98-r252-merged @ b42d6245d77fe30885ea8a90387771e1bc465e0f (SN120 reign-33 crown).
• Method: Offline DPO on teacher-anchored Reason pairs (chosen = higher lpC(y_C|z)−lpC(y_C|∅) thought). Optimized for Reason v3 (teacher-side only; no lpA / L1lift).
• Data: duel-derived Reason preference pairs (dpo_duel_reason.jsonl), ShortCtx HiRank filter; experiment mining/experiments/r576-r252-offline-dpo-hialpha-hirank-lobeta-shortctx-ultraextrasteps/.
• Hyperparameters: lr=5e-6, LoRA r=64 / α=128, β=0.02, max_len=6144, max_steps=2400 (stopped at 221 after data exhaustion / train.done), epoch=1.
• Hardware: train+merge+HF on mine-r252-vera-t4-nonking-grpo-1 (gentle-wolf-8c, 8×B300) GPUs 6,7.
• Axis note: ShortCtx×HiRank UltraExtraSteps on crowned r252. Distinct from ShortCtx×HiRank MidExtra R564 SIGNAL, ShortCtx MidRank Ultra R571, MidCtx Ultra R574, LongCtx Ultra R575, SoftCtx Ultra R573, Online R552, GRPO R565/R577.
• Sim evidence: n80 vs live king r252 reign-33 pending after HF push (SKIP_LOCAL_TKC; remote warm TKC). Decision rule: paired margin > max(2·SE, δ=0.002) and median thought ≥80 and B pass ≥0.30.
• Pass: Ralph p3466 merge+HF.