R1008 — MidCtx × HiRank × MidBeta Mega MidLR (offline DPO on vera king)

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 from dpo_duel_reason.jsonl under mining/experiments/r1008-vera-offline-dpo-hialpha-hirank-midbeta-midctx-megasuperextrasteps-ep4-midlr / pod /root/r1008/
• Key hyperparameters:
  • LoRA r=64 (HiRank), α=128 (HiAlpha)
  • β=0.1 (MidBeta)
  • lr=1e-6 (MidLR)
  • max_len=8192 (MidCtx)
  • max_steps=19200 (MegaSuperExtra)
  • epochs=4
• Hardware: Lium mine-r338-marsplan-online-dpo-bigg-hilr-1 (calm-fox-6a) 8×B200 GPUs 4,5 train+merge; TKC warm; chall :8003 GPUs 4,5 for v4 n80 → /tmp/r1008_merged (~16 safetensor shards)
• Local n80 vs live king reign36 (vera6/affine-5g4yy75zuz-t6@8e3f1695e058837ed80fec3238ff439fdc2d0f0e) under wvk=7:
  • margin +0.005917, SE 0.002233, z=2.650, n=80
  • bar max(2·SE, δ=0.002) = 0.004466 (~1.325×)
  • thought median 165 (≥80 ✓), B pass 0.433 (≥0.30 ✓)
  • k=3, τ=0.03 (fail-closed if stamp ≠ v4)
  • decision: WIN / Stage-5 licensed (r1008_sim_result_reign36_wvk7.json, p4150)
• Lineage: R1000 MidCtx HiRank Midβ Mega UltraLoLR REFUTE m=−0.000317 ~−0.15× → MidCtx HiRank Midβ Mega MidLR isolate between UltraLoLR R1000 and HiLR R991; ≠ Mega UltraLoLR R1000 / ≠ Mega HiLR R991 / ≠ SoftCtx Mega UltraLoLR R959 / ≠ SoftCtx HiRank Midβ Mega MidLR R992 / ≠ Online / ≠ GRPO
• Experiment path: mining/experiments/r1008-vera-offline-dpo-hialpha-hirank-midbeta-midctx-megasuperextrasteps-ep4-midlr

Intended use

License