Mosaic-IT: Cost-Free Compositional Data Synthesis for Instruction Tuning
This addresses the problem of expensive and non-sustainable data generation for instruction tuning in AI, though it is incremental as it builds on existing data synthesis techniques.
The paper tackles the high cost and limited diversity of instruction tuning for large language models by introducing Mosaic-IT, a method that synthesizes data by concatenating existing instructions, resulting in consistent performance improvements and an 80% reduction in training costs.
Finetuning large language models with a variety of instruction-response pairs has enhanced their capability to understand and follow instructions. Current instruction tuning primarily relies on teacher models or human intervention to generate and refine the instructions and responses for training, which are costly, non-sustainable, and may lack diversity. In this paper, we introduce Mosaic Instruction Tuning (Mosaic-IT), a human/model-free compositional data synthesis method that can efficiently create rich and diverse augmentations from existing instruction tuning data to enhance the LLMs. Mosaic-IT randomly concatenates multiple instruction data into one and trains the model to produce the corresponding responses with predefined higher-level meta-instructions to strengthen its multi-step instruction-following and format-following skills. Our extensive evaluations demonstrate a superior performance and training efficiency of Mosaic-IT, which achieves consistent performance improvements over various benchmarks and an 80% reduction in training costs compared with original instruction tuning. Our codes and data are available at https://github.com/tianyi-lab/Mosaic-IT.