Revisiting Catastrophic Forgetting in Large Language Model Tuning

TL;DR

This paper investigates the causes of catastrophic forgetting in large language models and proposes a sharpness-aware minimization technique to reduce it by flattening the loss landscape, demonstrating improved fine-tuning stability.

Contribution

It reveals the link between loss landscape flatness and catastrophic forgetting and introduces a novel mitigation method applicable across various LLMs.

Findings

Sharpness-aware minimization reduces CF in LLMs

Flattening the loss landscape improves fine-tuning stability

Method complements existing anti-forgetting strategies

Abstract

Catastrophic Forgetting (CF) means models forgetting previously acquired knowledge when learning new data. It compromises the effectiveness of large language models (LLMs) during fine-tuning, yet the underlying causes have not been thoroughly investigated. This paper takes the first step to reveal the direct link between the flatness of the model loss landscape and the extent of CF in the field of LLMs. Based on this, we introduce the sharpness-aware minimization to mitigate CF by flattening the loss landscape. Experiments on three widely-used fine-tuning datasets, spanning different model scales, demonstrate the effectiveness of our method in alleviating CF. Analyses show that we nicely complement the existing anti-forgetting strategies, further enhancing the resistance of LLMs to CF.