goal612/Affine-5czsc2fc98-r1032-vera-odpo-midrank-hibeta-shortctx-ultraextra-ep4-midlr-merged
The goal612/Affine-5czsc2fc98-r1032-vera-odpo-midrank-hibeta-shortctx-ultraextra-ep4-midlr-merged model is a 35.1 billion parameter language model derived from the vera6/affine-5g4yy75zuz-t6 base. It was trained using offline DPO on Reason-ranked duel pairs, specifically optimized for preference towards thoughts that enhance teacher-side Reason scores. With a context length of 32768 tokens, this model is intended for specialized use as an SN120 Affine miner submission and for evalsrv Reason v4 duels, rather than general chat applications.
Loading preview...
Overview
This model, goal612/Affine-5czsc2fc98-r1032-vera-odpo-midrank-hibeta-shortctx-ultraextra-ep4-midlr-merged, is a 35.1 billion parameter checkpoint developed by goal612. It is a challenger for Reason v4 (weight_version_key=7) within the Affine SN120 system, utilizing a tempered multi-sample log-mean-exp over k=3 teacher references.
Training Methodology
The model was trained via offline DPO (Direct Preference Optimization) on Reason-ranked duel pairs, not through Supervised Fine-Tuning (SFT) or online GRPO. The optimization goal was to instill a preference for thoughts that increase the teacher-side Reason score. Key training parameters include:
- LoRA: r=32 (MidRank), α=128 (HiAlpha)
- Beta: β=0.3 (HiBeta)
- Learning Rate: lr=1e-6 (MidLR)
- Context Length: max_len=6144 (ShortCtx) during training
- Epochs: 4
Performance and Validation
Against the live king vera6/affine-5g4yy75zuz-t6, this checkpoint achieved a margin of +0.005461 with a z-score of 2.098 over 80 samples, leading to a WIN / Stage-5 licensed decision. It passed the required thought median (175.5 ≥ 80) and B pass (0.475 ≥ 0.30) criteria.
Intended Use
This model is specifically designed for SN120 Affine miner submissions and evalsrv Reason v4 duels. It is not intended as a general-purpose chat model.