R596 โ€” SoftCtx ร— HiRank ร— MidBeta MegaExtra (offline DPO)

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

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 ร— HiRank pair set (soft context length band, high-rank filter); MidBeta ฮฒ=0.1. See experiment plan.md / dpo_duel_reason.jsonl under mining/experiments/r596-r252-offline-dpo-hialpha-hirank-midbeta-softctx-megaextrasteps.
  • Key hyperparameters:
    • LoRA r=64, ฮฑ=128
    • ฮฒ=0.1 (MidBeta)
    • lr=5e-6
    • max_len=12288 (SoftCtx)
    • max_steps target 3600 MegaExtra; TRAIN_DONE@259 (adapter kept / merged)
  • Hardware: Lium mine-r226-marsplan-fullft-1 (brave) GPUs 6,7 for train; merge + n80 on mine-r252-vera-t4-nonking-grpo-1 GPUs 4,5 โ†’ /tmp/r596_merged
  • Local n80 vs live king reign34 (cryptoDev23/Affine-5Dku3dYp9j-hk8161@55b7ffe0โ€ฆ):
    • margin +0.006196, SE 0.002357, z=2.63, n=75
    • bar max(2ยทSE, ฮด=0.002) = 0.004713 (~1.31ร—)
    • thought median 199 (โ‰ฅ80), B pass 0.368 (โ‰ฅ0.30)
    • decision: Stage-5 licensed (r596_decision_reign34.json)
  • Prior n80 vs r252: margin +0.008490 (~1.19ร— bar), thought/B clear
  • Experiment path: mining/experiments/r596-r252-offline-dpo-hialpha-hirank-midbeta-softctx-megaextrasteps

Intended use

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

License

Follows base model + Affine mining artifacts policy.

Downloads last month
13
Safetensors
Model size
35B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for elevateecho/sn120-3a778f1fd066

Finetuned
(18)
this model