YFC-112358/Qwen3.8-27B-TA-Aux-v1

Overview

YFC-112358/Qwen3.8-27B-TA-Aux-v1 is a 27 billion parameter model built upon the Qwen/Qwen3.8-27B base. Its core innovation lies in its construction method: it is an auxiliary model created through a specific "task arithmetic" merge of six distinct, weaker fine-tuned parent models. The merging process uses explicit coefficients, meaning there's no softmax or convex renormalization involved, and the c_strong coefficient is consistently 1.00.

Key Characteristics

• Task Arithmetic Merge: Combines six weaker fine-tuned models using a formula M_AUX = Qwen/Qwen3.8-27B + 3.0049 · mean_i(θ_i − Qwen/Qwen3.8-27B). This method pushes the displacement beyond the convex hull of the parent models.
• Explicit Coefficients: The merging coefficients are explicitly defined, not derived through normalization, offering precise control over the contribution of each parent.
• Lineage Tracking: The README provides detailed lineage, including commit SHAs for all parent models, ensuring reproducibility of the merging recipe.
• LoRA Integration: Some parent models are integrated as LoRA (Low-Rank Adaptation) sources, either as a base checkpoint with the adapter applied (compose) or as adapter-only contributions.

Intended Use

This model is explicitly designed as an auxiliary component ("辅料模型") to enhance or "season" stronger models, such as YFC-112358/Qwen3.8-27B-Seasoned-L020. It is not intended for standalone use, as it is likely to underperform its individual parent models when used in isolation. Its purpose is to act as a specialized additive to improve the performance or characteristics of a primary model, rather than serving as a general-purpose LLM itself. The merging process focuses on combining specific "weak" fine-tunes, and it does not include any "strong" models in its direct merge.