ShellFace/20260819-040753

TEXT GENERATIONConcurrent Unit Cost:3Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 19, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

ShellFace/20260819-040753 is a 35.1 billion parameter SN120 Affine challenger model, fine-tuned using offline DPO on Reason-ranked duel pairs. Developed by ShellFace, this model is optimized for improving 'Reason v4' scores in specific evaluation server duels. It is designed to enhance preference for thoughts that raise teacher-side Reason, making it suitable for specialized mining and evaluation tasks rather than general chat applications. The model utilizes a SoftCtx context length of 12288 tokens and was trained with LoRA (r=32, α=128) and a low learning rate (5e-7).

Loading preview...

ShellFace/20260819-040753: SN120 Affine Challenger

This model, developed by ShellFace, is an SN120 Affine challenger for Reason v4 (weight_version_key=7), a specific evaluation metric. It is a 35.1 billion parameter model fine-tuned using an offline DPO (Direct Preference Optimization) method, distinct from SFT or online GRPO. The training focused on optimizing preference for thoughts that increase teacher-side Reason scores, utilizing a tempered multi-sample log-mean-exp over k=3 teacher references.

Key Capabilities & Training Details

  • Base Model: vera6/affine-5g4yy75zuz-t6@8e3f1695e058837ed80fec3238ff439fdc2d0f0e.
  • Optimization Target: Enhancing Reason scores in duel pairs, specifically committing to a teacher next-action mode.
  • Data: Soft Mid Mid Soft × SoftCtx filtered duel preference pairs from dpo_duel_reason.jsonl.
  • Hyperparameters: Utilizes LoRA with r=32 (MidRank) and α=128 (HiAlpha), a β=0.1 (MidBeta), and an ultra-low learning rate (lr=5e-7).
  • Context Length: Supports a max_len=12288 (SoftCtx).
  • Performance: Achieved a margin of +0.003665 with a z-score of 2.177 against the live king reign36 under wvk=7, leading to a WIN / Stage-5 licensed decision.

Intended Use

This model is specifically designed for SN120 Affine miner submissions and evalsrv Reason v4 duels. It is not intended as a general chat model but rather for specialized evaluation and mining tasks where optimizing for the Reason v4 metric is critical.

Finetunes
1 models