ledgernova/sn120-9864cdf9071a

Model Overview

ledgernova/sn120-9864cdf9071a is a 35.1 billion parameter Affine SN120 challenger model, developed by ledgernova. It is specifically engineered for Reason v4 evaluations, employing a tempered multi-sample log-mean-exp over k=3 teacher references. The model's core mechanism involves calculating Reason = τ·log(mean_i exp(a_i/τ)) per turn, where a_i = lpC(y_i|z_A) − lpC(y_i|∅).

Training Methodology

This checkpoint was trained using offline DPO (Direct Preference Optimization) on Reason-ranked duel pairs, distinct from SFT or online GRPO methods. The optimization focused on enhancing preference for thoughts that increase teacher-side Reason. Key hyperparameters during training included a LoRA rank of 32 (MidRank), α=128 (HiAlpha), β=0.1 (MidBeta), and an ultra-low learning rate of 5e-7 (UltraLoLR). It was trained for 4 epochs with a maximum context length of 12288 tokens.

Performance and Intended Use

During its evaluation against the live king vera6/affine-5g4yy75zuz-t6 under wvk=7, this model achieved a margin of +0.003665 with a z-score of 2.177 over 80 samples, leading to its WIN / Stage-5 licensed decision. The model's thought median was 141.5 and B pass was 0.5375. It is explicitly intended for SN120 Affine miner submissions and evalsrv Reason v4 duels, and not designed as a general-purpose chat model.