IC-299Instruction tuning preserves the weight-matrix structure of LLMs, with DOCS scores exceeding 0.7 across all matrices

Zeping Min, Xinshang Wang

SourceDOCS: Quantifying Weight Similarity for Deeper Insights into Large Language Models

The paper compares weight matrices of base and instruction-tuned models in the Yi-1.5, Llama-3.1, and Gemma-2 families using the DOCS index. All DOCS scores between corresponding layers of the base and instruct variants exceed 0.7, indicating that instruction fine-tuning largely retains the foundational weight structure. Additionally, the weight matrices cluster into three functional groups based on their DOCS score trends: MLP-up with MLP-down, Wq with Wk, and Wv with Wo.

Evidence
correlational
Key metric
all docs scores are notably high, with values exceeding 0.7 for every matrix evaluated
Model
Yi Yi-1.5-9B, Llama 3.1 8B, 8B Instruct, Gemma 2 9B
Related work
Ouyang et al. 2022 (instruction tuning) [context]
Related findings
IC-298, IC-300, IC-301
Extraction
automatic-extraction