crazyape777/fk-unconst-affine-5czsc2fc98-r1064-vera-odpo-midrank-midbeta-mi
The crazyape777/fk-unconst-affine-5czsc2fc98-r1064-vera-odpo-midrank-midbeta-mi is a 35.1 billion parameter Affine SN120 challenger model, fine-tuned using offline DPO on Reason-ranked duel pairs. It is specifically optimized for the Reason v4 evaluation metric, demonstrating a preference for thoughts that enhance teacher-side Reason scores. This model is intended for specialized use in Affine miner submissions and evalsrv Reason v4 duels, rather than general chat applications.
Loading preview...
Model Overview
This model, fk-unconst-affine-5czsc2fc98-r1064-vera-odpo-midrank-midbeta-mi, is a 35.1 billion parameter Affine SN120 challenger, derived from the vera6/affine-5g4yy75zuz-t6 base model. It was trained using an offline DPO (Direct Preference Optimization) method, specifically on Reason-ranked duel pairs, rather than traditional SFT or online GRPO.
Key Characteristics & Training
- Optimization Target: The model was optimized to show a strong preference for responses that increase the teacher-side Reason score, utilizing a tempered multi-sample log-mean-exp over three teacher references.
- Training Data: Fine-tuned on a dataset of 604 "Soft Mid Mid Soft → MidCtx filtered duel preference pairs" (
dpo_duel_reason.jsonl). - Hyperparameters: Notable hyperparameters include a LoRA rank of 32 (MidRank), alpha of 128 (HiAlpha), beta of 0.1 (MidBeta), a learning rate of 2e-6 (HiLR), and a maximum context length of 8192 tokens (MidCtx).
- Performance: Achieved a margin of +0.006632 with a z-score of 2.042 against the live king
reign36underwvk=7, leading to a "WIN / Stage-5 licensed" decision.
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.