unc-ffb22875 API & Inference Endpoint

How this checkpoint was trained

Base / parent: unconst/Affine-5czsc2fc98-r252-merged@b42d6245d77fe30885ea8a90387771e1bc465e0f (our crowned r252, reign 33)
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: ShortCtx × MidRank × HiBeta filtered duel preference pairs from dpo_duel_reason.jsonl under mining/experiments/r683-r252-offline-dpo-hialpha-midrank-hibeta-shortctx-ultraextrasteps-ep3-lolr (kept ≥200 examples at launch)
Key hyperparameters:
- LoRA r=32 (MidRank), α=128 (HiAlpha)
- β=0.3 (HiBeta)
- lr=1e-6 (LoLR)
- max_len=6144 (ShortCtx)
- max_steps=7200 (UltraExtra; 2× Mega 3600)
- epochs=3
Hardware: Lium mine-r226-marsplan-fullft-1 (brave-raven-a9) 8×B200 GPUs 6,7 train+merge; SCP → mine-r262-kevin-v5-nonking-grpo-1 (golden-comet-78) GPUs 4,5 challenger serve + v4 n80 → /tmp/r683_merged (~66G / 16 safetensor shards)
Local n80 vs live king reign34 (cryptoDev23/Affine-5Dku3dYp9j-hk8161@55b7ffe0…) under wvk=7:
- margin +0.002137, SE 0.000943, z=2.27, n=79
- bar max(2·SE, δ=0.002) = 0.002 (~1.07×)
- thought median 172 (≥80 ✓), B pass 0.304 (≥0.30 ✓, knife-edge)
- k=3, τ=0.03 (fail-closed if stamp ≠ v4)
- decision: WIN / Stage-5 licensed (r683_decision_reign34_wvk7.json, p3738)
Lineage: amplify R659 Short MidRank HiBeta Mega / R622 ~0.85× with 2× steps; ≠ R679 Short HiRank HiBeta UltraExtra / ≠ R653 Short MidRank LoBeta / ≠ R658 Short MidRank MidBeta / ≠ Online / ≠ GRPO
Experiment path: mining/experiments/r683-r252-offline-dpo-hialpha-midrank-hibeta-shortctx-ultraextrasteps-ep3-lolr

Intended use

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

License

Follows base model + Affine mining artifacts policy.

How this checkpoint was trained

Base / parent: unconst/Affine-5czsc2fc98-r252-merged@b42d6245d77fe30885ea8a90387771e1bc465e0f (our crowned r252, reign 33)
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: ShortCtx × MidRank × HiBeta filtered duel preference pairs from dpo_duel_reason.jsonl under mining/experiments/r683-r252-offline-dpo-hialpha-midrank-hibeta-shortctx-ultraextrasteps-ep3-lolr (kept ≥200 examples at launch)
Key hyperparameters:
- LoRA r=32 (MidRank), α=128 (HiAlpha)
- β=0.3 (HiBeta)
- lr=1e-6 (LoLR)
- max_len=6144 (ShortCtx)
- max_steps=7200 (UltraExtra; 2× Mega 3600)
- epochs=3
Hardware: Lium mine-r226-marsplan-fullft-1 (brave-raven-a9) 8×B200 GPUs 6,7 train+merge; SCP → mine-r262-kevin-v5-nonking-grpo-1 (golden-comet-78) GPUs 4,5 challenger serve + v4 n80 → /tmp/r683_merged (~66G / 16 safetensor shards)
Local n80 vs live king reign34 (cryptoDev23/Affine-5Dku3dYp9j-hk8161@55b7ffe0…) under wvk=7:
- margin +0.002137, SE 0.000943, z=2.27, n=79
- bar max(2·SE, δ=0.002) = 0.002 (~1.07×)
- thought median 172 (≥80 ✓), B pass 0.304 (≥0.30 ✓, knife-edge)
- k=3, τ=0.03 (fail-closed if stamp ≠ v4)
- decision: WIN / Stage-5 licensed (r683_decision_reign34_wvk7.json, p3738)
Lineage: amplify R659 Short MidRank HiBeta Mega / R622 ~0.85× with 2× steps; ≠ R679 Short HiRank HiBeta UltraExtra / ≠ R653 Short MidRank LoBeta / ≠ R658 Short MidRank MidBeta / ≠ Online / ≠ GRPO
Experiment path: mining/experiments/r683-r252-offline-dpo-hialpha-midrank-hibeta-shortctx-ultraextrasteps-ep3-lolr

Intended use

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

License

Follows base model + Affine mining artifacts policy.

unc-ffb22875

README

How this checkpoint was trained

Intended use

License

Explore FriendliAI today

README

How this checkpoint was trained

Intended use

License