Supervised fine-tuning (SFT) is a critical step in aligning large language models (LLMs) with human instructions and values, yet many aspects of SFT remain poorly understood. We trained a wide range of base models on a variety of datasets including code generation, mathematical reasoning, and general-domain tasks, resulting in 1,000+ SFT models under controlled conditions. We then identified the dataset properties that matter most and examined the layer-wise modifications introduced by SFT. Our findings reveal that some training-task synergies persist across all models while others vary substantially, emphasizing the importance of model-specific strategies. Moreover, we demonstrate that perplexity consistently predicts SFT effectiveness, often surpassing superficial similarity between the training data and the benchmark, and that mid-layer weight changes correlate most strongly with performance gains. We release these 1,000+ SFT models and benchmark results to accelerate further research. All resources are available at https://github.com/llm-jp/massive-sft.
@article{arxiv.2506.14681,
title = {Massive Supervised Fine-tuning Experiments Reveal How Data, Layer, and Training Factors Shape LLM Alignment Quality},
author = {Yuto Harada and Yusuke Yamauchi and Yusuke Oda and Yohei Oseki and Yusuke Miyao and Yu Takagi},
journal= {arXiv preprint arXiv:2506.14681},
year = {2025}
}
Comments
Accepted to EMNLP 2025 (Main Conference). Models and evaluation results available at: https://github.com/llm-jp/massive-sft