ledgernova/sn120-ff9153704c9d
The ledgernova/sn120-ff9153704c9d is a 35.1 billion parameter Affine SN120 model, developed by ledgernova, specifically fine-tuned using offline DPO on Reason-ranked duel pairs. It is optimized to prefer thoughts that enhance teacher-side Reason scores, utilizing a tempered multi-sample log-mean-exp over teacher references. This model is intended for SN120 Affine miner submissions and evaluation server Reason v4 duels, rather than general chat applications.
Loading preview...
Model Overview
The ledgernova/sn120-ff9153704c9d is a 35.1 billion parameter Affine SN120 model, developed by ledgernova, designed as a challenger for Reason v4. It is built upon the vera6/affine-5g4yy75zuz-t6 base model.
Key Characteristics and Training
This model was trained using offline DPO (Direct Preference Optimization) on Reason-ranked duel pairs, rather than Supervised Fine-Tuning (SFT) or online GRPO. The primary optimization goal was to foster a preference for thoughts that elevate the teacher-side Reason score, which is calculated using a tempered multi-sample log-mean-exp over three teacher references. Key training hyperparameters include a LoRA rank of 64 (HiRank), alpha of 128 (HiAlpha), beta of 0.1 (MidBeta), and a learning rate of 1e-6 (MidLR). It was trained for 19200 steps across 4 epochs with a maximum context length of 8192 tokens.
Performance and Intended Use
During its development, the model demonstrated a positive margin of +0.005917 against the live king vera6/affine-5g4yy75zuz-t6 under wvk=7, achieving a z-score of 2.650 over 80 samples. It passed the required thought median (165, ≥80) and B pass (0.433, ≥0.30) criteria, leading to a WIN / Stage-5 licensed decision. This model is specifically intended for SN120 Affine miner submissions and evaluation server Reason v4 duels, and is not designed as a general-purpose chat model.