windsword8989/5GbZvZ7tcC-d1

R861 — SoftCtx × MidRank × MidBeta UltraLoLR (offline DPO on vera king)

Affine SN120 challenger for Reason v4 (weight_version_key=7): tempered multi-sample log-mean-exp over k=3 teacher refs (τ=0.03).

Per turn: a_i = lpC(y_i|z_A) − lpC(y_i|∅); Reason = τ·log(mean_i exp(a_i/τ)). Crown also needs median stripped |z|≥80 and B pass ≥0.30.

How this checkpoint was trained

• —Base / parent: vera6/affine-5g4yy75zuz-t6@8e3f1695e058837ed80fec3238ff439fdc2d0f0e (live king reign36)
• —Method: offline DPO on Reason-ranked duel pairs (not SFT / not online GRPO)
• —What was optimized: preference for thoughts that raise teacher-side Reason (commit to a teacher next-action mode; filler loses under LME)
• —Data: Soft Mid Mid Soft × SoftCtx filtered duel preference pairs from dpo_duel_reason.jsonl under mining/experiments/r861-vera-offline-dpo-hialpha-midrank-midbeta-softctx-megasuperextrasteps-ep4-ultralolr / pod /root/r861/ (~259–604 rows at launch)
• —Key hyperparameters:
• —LoRA r=32 (MidRank), α=128 (HiAlpha)
• —β=0.1 (MidBeta)
• —lr=5e-7 (UltraLoLR)
• —max_len=12288 (SoftCtx)
• —max_steps=19200 (MegaSuperExtra)
• —epochs=4
• —Hardware: Lium mine-crown-1 (gentle-orbit-bd) 8×B200 GPUs 4,5 train+merge+challenger serve + v4 n80 → /tmp/r861_merged (~66G / 16 safetensor shards; weight_identical=false)
• —Local n80 vs live king reign36 (vera6/affine-5g4yy75zuz-t6@8e3f1695e058837ed80fec3238ff439fdc2d0f0e) under wvk=7:
• —margin +0.003665, SE 0.001684, z=2.177, n=80
• —bar max(2·SE, δ=0.002) = 0.003367 (~1.088×)
• —thought median 141.5 (≥80 ✓), B pass 0.5375 (≥0.30 ✓)
• —k=3, τ=0.03 (fail-closed if stamp ≠ v4)
• —decision: WIN / Stage-5 licensed (r861_decision_reign36_wvk7.json, p3966)
• —Lineage: R846 vera MidRank HiBeta SoftCtx REFUTE ~−0.33× → MidBeta isolate; ≠ R846 Hiβ / ≠ R847 HiRank Midβ SoftCtx / ≠ Online / ≠ GRPO
• —Experiment path: mining/experiments/r861-vera-offline-dpo-hialpha-midrank-midbeta-softctx-megasuperextrasteps-ep4-ultralolr

Intended use

SN120 Affine miner submission / evalsrv Reason v4 duel. Not a general chat model.

License

Follows base model + Affine mining artifacts policy.