unconst/Affine-5czsc2fc98-r1064-vera-odpo-midrank-midbeta-midctx-ultraextra-ep4-hilr-merged

R1064 — MidCtx × MidRank × MidBeta Ultra HiLR (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 → MidCtx filtered duel preference pairs (dpo_duel_reason.jsonl, 604 lines) under mining/experiments/r1064-vera-offline-dpo-hialpha-midrank-midbeta-midctx-ultrasuperextrasteps-ep4-hilr / pod /root/r1064/
• —Key hyperparameters:
• —LoRA r=32 (MidRank), α=128 (HiAlpha)
• —β=0.1 (MidBeta)
• —lr=2e-6 (HiLR)
• —max_len=8192 (MidCtx)
• —max_steps=28800 (UltraSuperExtra)
• —epochs=4
• —Hardware: Lium mine-r337-marsplan-online-dpo-hilr-1 (noble-hawk-1f) 8×B200 GPUs 4,5 train+merge; TKC warm; chall :8003 GPUs 4,5 for v4 n80 → /tmp/r1064_merged (~16 safetensor shards)
• —Local n80 vs live king reign36 (vera6/affine-5g4yy75zuz-t6@8e3f1695e058837ed80fec3238ff439fdc2d0f0e) under wvk=7:
• —margin +0.006632, SE 0.003248, z=2.042, n=79
• —bar max(2·SE, δ=0.002) = 0.006495 (~1.021×)
• —thought median 201 (≥80 ✓), B pass 0.521 (≥0.30 ✓)
• —k=3, τ=0.03 (fail-closed if stamp ≠ v4)
• —decision: WIN / Stage-5 licensed (r1064_sim_result_reign36_wvk7.json, p4208)
• —Lineage: R1047 MidCtx MidRank MidLoβ Ultra HiLR REFUTE m=+0.002013 ~0.19× → Midβ isolate (MidLoβ→Midβ); ≠ MidLoβ Ultra HiLR R1047 / ≠ MidCtx MidRank Hiβ Ultra HiLR R1053 / ≠ MidCtx LoRank Midβ Ultra HiLR R1057 / ≠ ShortCtx MidRank Midβ Ultra HiLR R1063 / ≠ Online / ≠ GRPO
• —Experiment path: mining/experiments/r1064-vera-offline-dpo-hialpha-midrank-midbeta-midctx-ultrasuperextrasteps-ep4-hilr

Intended use

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

License

Follows base model + Affine mining artifacts policy.