Dynamic Data Mixing Maximizes Instruction Tuning for Mixture-of-Experts
This work addresses the challenge of efficiently training MoE models on multiple tasks for researchers and practitioners in AI, representing an incremental improvement over fixed sampling methods.
The paper tackles the problem of suboptimal performance in Mixture-of-Experts (MoE) models during instruction tuning by proposing a dynamic data mixing method that adjusts dataset sampling weights based on inter-redundancies, resulting in improved performance on downstream knowledge and reasoning tasks and open-ended queries.
Mixture-of-Experts (MoE) models have shown remarkable capability in instruction tuning, especially when the number of tasks scales. However, previous methods simply merge all training tasks (e.g. creative writing, coding, and mathematics) and apply fixed sampling weights, without considering the importance of different tasks as the model training state changes. In this way, the most helpful data cannot be effectively distinguished, leading to suboptimal model performance. To reduce the potential redundancies of datasets, we make the first attempt and propose a novel dynamic data mixture for MoE instruction tuning. Specifically, inspired by MoE's token routing preference, we build dataset-level representations and then capture the subtle differences among datasets. Finally, we propose to dynamically adjust the sampling weight of datasets by their inter-redundancies, thus maximizing global performance under a limited training budget. The experimental results on two MoE models demonstrate the effectiveness of our approach on both downstream knowledge \& reasoning tasks and open-ended queries. Code and models are available at https://github.com/Spico197/MoE-SFT .