MoFO: Momentum-Filtered Optimizer for Mitigating Forgetting in LLM Fine-Tuning

TL;DR

MoFO is a novel optimizer for fine-tuning large language models that selectively updates parameters with high momentum to reduce knowledge forgetting without requiring access to pre-training data.

Contribution

MoFO introduces a momentum-based selective update algorithm that mitigates forgetting during fine-tuning of LLMs without needing pre-training data.

Findings

MoFO achieves comparable performance to standard fine-tuning methods.

MoFO effectively reduces knowledge forgetting in LLMs.

Theoretical convergence guarantees support MoFO's effectiveness.

Abstract

Large language models (LLMs) have demonstrated remarkable capabilities across a wide range of tasks. Typically, LLMs are first pre-trained on large corpora and subsequently fine-tuned on task-specific datasets. However, during fine-tuning, LLMs may forget some knowledge acquired in the pre-training stage, leading to a decline in general capabilities. Existing approaches to mitigate forgetting often rely on access to pre-training data, which may be unavailable in many real-world scenarios--such as fine-tuning checkpoint-only open-source LLMs. To address this challenge, we propose a new fine-tuning algorithm termed Momentum-Filtered Optimizer (MoFO). MoFO is an extension of greedy block coordinate descent (BCD) methods: in each iteration, MoFO only updates the model parameters with the largest momentum magnitudes, while keeping all other parameters fixed. MoFO achieves similar fine-tuning performance to the default fine-tuning algorithm while effectively mitigating knowledge forgetting. We validate MoFO through rigorous convergence analysis and extensive experiments, demonstrating its effectiveness in mitigating forgetting without pre-training data.