@inproceedings{pham-etal-2025-fine,
    title = "How to Fine-Tune Safely on a Budget: Model Adaptation Using Minimal Resources",
    author = "Pham, Anh C.  and
      Thalanki, Mihir  and
      Sun, Michael  and
      Chaloo, Aditya  and
      Gupta, Ankita  and
      Xia, Tian  and
      Mate, Aditya  and
      Nosakhare, Ehi  and
      Srinivasan, Soundararajan",
    editor = "Potdar, Saloni  and
      Rojas-Barahona, Lina  and
      Montella, Sebastien",
    booktitle = "Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track",
    month = nov,
    year = "2025",
    address = "Suzhou (China)",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/ingest-emnlp/2025.emnlp-industry.138/",
    pages = "1970--1981",
    ISBN = "979-8-89176-333-3",
    abstract = "Supervised fine-tuning (SFT) on benign data can paradoxically erode a language model{'}s safety alignment, a phenomenon known as catastrophic forgetting of safety behaviors. Although prior work shows that randomly adding safety examples can reduce harmful output, the principles that make certain examples more effective than others remain poorly understood. This paper investigates the hypothesis that the effectiveness of a safety example is governed by two key factors: its instruction-response behavior (e.g., refusal vs. explanation) and its semantic diversity across harm categories. We systematically evaluate sampling strategies based on these axes and find that structured, diversity-aware sampling significantly improves model safety. Our method reduces harmfulness by up to 41{\%} while adding only 0.05{\%} more data to the fine-tuning set."
}