Affine-5czsc2fc98-r565-r252-king-grpo-hialpha-merged

Training story

• Base / parent king: unconst/Affine-5czsc2fc98-r252-merged @ b42d6245d77fe30885ea8a90387771e1bc465e0f (SN120 reign-33 crown).
• Method: GRPO on teacher-side Reason reward (group rollouts scored by lpC(y_C|z)−lpC(y_C|∅)). Optimized for Reason v3 (teacher-side only; no lpA / L1lift).
• Data: high-Reason winner thoughts (winner_za_high_l1.jsonl / GRPO online samples); experiment mining/experiments/r565-r252-king-grpo-hialpha/.
• Hyperparameters: lr=5e-6, LoRA r=16 / α=128, group_size=4, max_len=6144, max_steps=200 (stopped at 151), temp=0.8.
• Hardware: train+merge+HF on mine-r252-vera-t4-nonking-grpo-1 (gentle-wolf-8c, 8×B300) GPUs 4,5.
• Axis note: R252 crown GRPO recipe reapplied to crowned r252 weights (king-base HiAlpha-GRPO). Distinct from Offline UltraExtra R570/R571/R572 and Online R552.
• 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 p3449 merge+HF.