crazyape777/fk-unconst-affine-5czsc2fc98-r938-vera-odpo-midrank-hibeta-soft
The crazyape777/fk-unconst-affine-5czsc2fc98-r938-vera-odpo-midrank-hibeta-soft model is a 35.1 billion parameter Affine SN120 challenger, fine-tuned using offline DPO on Reason-ranked duel pairs. It is optimized for improving preference for thoughts that raise teacher-side Reason, rather than general chat. This model features a 32768 token context length and utilizes LoRA with r=32 and α=128.
Loading preview...
Model Overview
The crazyape777/fk-unconst-affine-5czsc2fc98-r938-vera-odpo-midrank-hibeta-soft model is a 35.1 billion parameter Affine SN120 challenger, specifically developed for Reason v4 (weight_version_key=7). It is based on the vera6/affine-5g4yy75zuz-t6 parent model and was trained using an offline DPO (Direct Preference Optimization) method, not Supervised Fine-Tuning (SFT) or online GRPO.
Key Characteristics
- Optimization Target: The model was optimized to enhance preference for thoughts that increase teacher-side Reason, using a tempered multi-sample log-mean-exp over k=3 teacher references.
- Training Data: Utilized Soft Mid Mid Soft × SoftCtx filtered duel preference pairs from
dpo_duel_reason.jsonl. - Hyperparameters: Notable settings include LoRA with
r=32(MidRank) andα=128(HiAlpha),β=0.3(HiBeta), a very low learning rate (lr=5e-7, UltraLoLR), and a substantial maximum context length of12288tokens (SoftCtx). - Performance: Achieved a win against the live king
reign36with a margin of+0.004951and a z-score of2.399on 80 samples, meeting the criteria for Stage-5 licensing.
Intended Use
This model is specifically designed as an SN120 Affine miner submission and for evalsrv Reason v4 duels. It is not intended as a general-purpose chat model.