elevateecho/sn120-7aded176a1f1
The elevateecho/sn120-7aded176a1f1 is a 35.1 billion parameter Affine SN120 challenger model, fine-tuned using offline DPO on Reason-ranked pairs. It is specifically optimized for reasoning tasks, aiming to surpass previous models in teacher-anchored Reason scores. This model is intended for specialized SN120 Affine miner submissions and evaluation server Reason duels, rather than general chat applications.
Loading preview...
Model Overview
The elevateecho/sn120-7aded176a1f1 is a 35.1 billion parameter Affine SN120 challenger model, developed by unconst/Affine-5czsc2fc98-r252-merged. It was trained using an offline DPO (Direct Preference Optimization) method, specifically on pairs ranked by a 'Reason' metric, rather than through Supervised Fine-Tuning (SFT) or online GRPO.
Key Training Details
- Optimization Goal: Enhanced preference for higher teacher-side Reason scores on mined data pairs.
- Data Set: Utilized a
SoftCtx × MidRank × LoBetapair set, characterized by a soft context band, mid LoRA rank, and low DPO β. - Hyperparameters: Notable settings include LoRA r=32 (MidRank), α=128 (HiAlpha), a low β=0.02 (LoBeta), and a learning rate of 1e-6 (LoLR).
- Context Length: Trained with a maximum sequence length of 12288 tokens (SoftCtx).
- Performance: This checkpoint demonstrated a positive margin of +0.005735 against the live king reign34 model (
cryptoDev23/Affine-5Dku3dYp9j-hk8161), with a z-score of 2.91 over 77 samples, indicating a statistically significant improvement in reasoning capabilities.
Intended Use
This model is specifically designed for SN120 Affine miner submissions and evaluation server Reason duels. It is not intended for use as a general-purpose chat model.