michael-chan-000/affine-5GbZvZ7tcC-d1
The michael-chan-000/affine-5GbZvZ7tcC-d1 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 improving 'Reason v4' performance in evaluation server duels, focusing on preference for thoughts that raise teacher-side Reason. This model is not intended as a general chat model but rather for specialized mining submissions and evaluations.
Loading preview...
Model Overview
michael-chan-000/affine-5GbZvZ7tcC-d1 is a 35.1 billion parameter model, an Affine SN120 challenger for Reason v4. It was developed by michael-chan-000 and is specifically designed for specialized evaluation server duels, not as a general-purpose chat model.
Training Methodology
This checkpoint was trained using offline DPO (Direct Preference Optimization) on Reason-ranked duel pairs, diverging from traditional SFT (Supervised Fine-Tuning) or online GRPO methods. The optimization focused on enhancing preference for thoughts that elevate teacher-side Reason, utilizing a tempered multi-sample log-mean-exp over k=3 teacher references.
Key Training Details:
- Base Model:
vera6/affine-5g4yy75zuz-t6@8e3f1695e058837ed80fec3238ff439fdc2d0f0e - Data: Soft Mid Mid Soft × SoftCtx filtered duel preference pairs from
dpo_duel_reason.jsonl. - Hyperparameters: Notable settings include LoRA r=32 (MidRank), α=128 (HiAlpha), β=0.1 (MidBeta), and a learning rate of 5e-7 (UltraLoLR).
- Context Length: Trained with a maximum sequence length of 12288 tokens (SoftCtx).
Performance & Intended Use
The model demonstrated a positive margin of +0.003665 against its predecessor in local evaluations, leading to its licensing as a Stage-5 WIN. Its primary and sole intended use is for SN120 Affine miner submissions and evalsrv Reason v4 duels. It is explicitly not designed or recommended as a general chat model.