sn120-7aded176a1f1

R637 — SoftCtx × MidRank × LoBeta MegaExtra ep3 × LoLR (offline DPO)

Affine SN120 challenger trained to beat the live king on Reason v3 (teacher-anchored score: lpC(y_C|z_A) − lpC(y_C|∅)).

How this checkpoint was trained

• Base / parent: unconst/Affine-5czsc2fc98-r252-merged@b42d6245d77fe30885ea8a90387771e1bc465e0f (our crowned r252, reign 33)
• Method: offline DPO on Reason-ranked pairs (not SFT / not online GRPO)
• What was optimized: preference for higher teacher-side Reason on mined pairs
• Data: SoftCtx × MidRank × LoBeta pair set (soft context band, mid LoRA rank, low DPO β). See plan.md / dpo_duel_reason.jsonl under mining/experiments/r637-r252-offline-dpo-hialpha-midrank-lobeta-softctx-megaextrasteps-ep3-lolr.
• Key hyperparameters:
  • LoRA r=32 (MidRank), α=128 (HiAlpha)
  • β=0.02 (LoBeta)
  • lr=1e-6 (LoLR)
  • max_len=12288 (SoftCtx)
  • max_steps=3600 MegaExtra; epochs=3
• Hardware: Lium mine-r226-marsplan-fullft-1 (brave) GPUs 2,3 train+merge; host-relay → n80 on mine-r262-kevin-v5-nonking-grpo-1 (golden) GPUs 6,7 → /tmp/r637_merged
• Local n80 vs live king reign34 (cryptoDev23/Affine-5Dku3dYp9j-hk8161@55b7ffe0…):
  • margin +0.005735, SE 0.001973, z=2.91, n=77
  • bar max(2·SE, δ=0.002) = 0.003946 (~1.45×)
  • thought median 168.5 (≥80), B pass 0.4125 (≥0.30)
  • decision: Stage-5 licensed (r637_decision_reign34.json, p3643)
• Lineage: amplify of R580 Soft MidRank LoBeta Mega ep1 ~0.39× after R610 ep2 @5e-6 REFUTE; ≠ Soft MidRank MidBeta R635 ep3 LoLR REFUTE
• Experiment path: mining/experiments/r637-r252-offline-dpo-hialpha-midrank-lobeta-softctx-megaextrasteps-ep3-lolr

Intended use

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

License

Follows base model + Affine mining artifacts policy.