R1032 โ€” ShortCtx ร— MidRank ร— HiBeta Ultra MidLR (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 (ฯ„=0.03).

Per turn: a_i = lpC(y_i|z_A) โˆ’ lpC(y_i|โˆ…); Reason = ฯ„ยทlog(mean_i exp(a_i/ฯ„)). Crown also needs median stripped |z|โ‰ฅ80 and B pass โ‰ฅ0.30.

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 โ†’ ShortCtx filtered duel preference pairs (dpo_duel_reason.jsonl, 604 lines) under mining/experiments/r1032-vera-offline-dpo-hialpha-midrank-hibeta-shortctx-ultrasuperextrasteps-ep4-midlr / pod /root/r1032/
  • Key hyperparameters:
    • LoRA r=32 (MidRank), ฮฑ=128 (HiAlpha)
    • ฮฒ=0.3 (HiBeta)
    • lr=1e-6 (MidLR)
    • max_len=6144 (ShortCtx)
    • max_steps=28800 (UltraSuperExtra)
    • epochs=4
  • Hardware: Lium mine-r924-vera-midctx-hibeta-1 (cosmic-orbit-55) 8ร—H200 GPUs 1,3 train+merge; TKC warm; chall :8003 GPUs 1,3 for v4 n80 โ†’ /tmp/r1032_merged (~16 safetensor shards)
  • Local n80 vs live king reign36 (vera6/affine-5g4yy75zuz-t6@8e3f1695e058837ed80fec3238ff439fdc2d0f0e) under wvk=7:
    • margin +0.005461, SE 0.002603, z=2.098, n=80
    • bar max(2ยทSE, ฮด=0.002) = 0.005206 (~1.049ร—)
    • thought median 175.5 (โ‰ฅ80 โœ“), B pass 0.475 (โ‰ฅ0.30 โœ“)
    • k=3, ฯ„=0.03 (fail-closed if stamp โ‰  v4)
    • decision: WIN / Stage-5 licensed (r1032_sim_result_reign36_wvk7.json, p4177)
  • Lineage: R1015 ShortCtx MidRank Hiฮฒ Mega MidLR REFUTE m=โˆ’0.002353 ~โˆ’0.94ร— โ†’ ShortCtx MidRank Hiฮฒ Ultra MidLR isolate (Megaโ†’Ultra); โ‰  Mega MidLR R1015 / โ‰  Mega UltraLoLR R1001 / โ‰  Ultra HiLR R984 / โ‰  SoftCtx MidRank Hiฮฒ Mega MidLR R1027 / โ‰  MidCtx MidRank Hiฮฒ Ultra MidLR R1017 / โ‰  Online / โ‰  GRPO
  • Experiment path: mining/experiments/r1032-vera-offline-dpo-hialpha-midrank-hibeta-shortctx-ultrasuperextrasteps-ep4-midlr

Intended use

SN120 Affine miner submission / evalsrv Reason v4 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 tastegarden/sn120-62dfb322fdce

Finetuned
(35)
this model