Preventing Catastrophic Forgetting: Behavior-Aware Sampling for Safer Language Model Fine-Tuning

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Pham, Anh, Thalanki, Mihir, Sun, Michael, Chaloo, Aditya, Gupta, Ankita, Xia, Tian, Mate, Aditya, Nosakhare, Ehimwenma, Srinivasan, Soundararajan
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914114040758272
author Pham, Anh
Thalanki, Mihir
Sun, Michael
Chaloo, Aditya
Gupta, Ankita
Xia, Tian
Mate, Aditya
Nosakhare, Ehimwenma
Srinivasan, Soundararajan
author_facet Pham, Anh
Thalanki, Mihir
Sun, Michael
Chaloo, Aditya
Gupta, Ankita
Xia, Tian
Mate, Aditya
Nosakhare, Ehimwenma
Srinivasan, Soundararajan
contents Large language models often lose previously aligned safety behaviors when fine-tuned on benign data, a phenomenon known as catastrophic forgetting. Prior work shows that adding random safety examples can mitigate this effect, but it remains unclear which examples are most effective. We propose a behavior-aware sampling framework that selects safety examples based on two complementary factors: instruction-response behavior (e.g., refusal versus compliance) and semantic diversity across harm categories. Systematic evaluation shows that this approach substantially reduces harmful outputs while maintaining helpfulness, achieving up to a 41% reduction in harmfulness with only 0.5% additional training data. These results highlight how targeted data selection can improve the safety and efficiency of fine-tuning at scale.
format Preprint
id arxiv_https___arxiv_org_abs_2510_21885
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Preventing Catastrophic Forgetting: Behavior-Aware Sampling for Safer Language Model Fine-Tuning
Pham, Anh
Thalanki, Mihir
Sun, Michael
Chaloo, Aditya
Gupta, Ankita
Xia, Tian
Mate, Aditya
Nosakhare, Ehimwenma
Srinivasan, Soundararajan
Computation and Language
Artificial Intelligence
Large language models often lose previously aligned safety behaviors when fine-tuned on benign data, a phenomenon known as catastrophic forgetting. Prior work shows that adding random safety examples can mitigate this effect, but it remains unclear which examples are most effective. We propose a behavior-aware sampling framework that selects safety examples based on two complementary factors: instruction-response behavior (e.g., refusal versus compliance) and semantic diversity across harm categories. Systematic evaluation shows that this approach substantially reduces harmful outputs while maintaining helpfulness, achieving up to a 41% reduction in harmfulness with only 0.5% additional training data. These results highlight how targeted data selection can improve the safety and efficiency of fine-tuning at scale.
title Preventing Catastrophic Forgetting: Behavior-Aware Sampling for Safer Language Model Fine-Tuning
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2510.21885