Star-Agents: Automatic Data Optimization with LLM Agents for Instruction Tuning

TL;DR

A novel Star-Agents framework is proposed, which automates the enhancement of data quality across datasets through multi-agent collaboration and assessment, and evolves in a dynamic refinement phase, where more effective LLMs are prioritized, enhancing the overall data quality.

Abstract

The efficacy of large language models (LLMs) on downstream tasks usually hinges on instruction tuning, which relies critically on the quality of training data. Unfortunately, collecting high-quality and diverse data is both expensive and time-consuming. To mitigate this issue, we propose a novel Star-Agents framework, which automates the enhancement of data quality across datasets through multi-agent collaboration and assessment. The framework adopts a three-pronged strategy. It initially generates diverse instruction data with multiple LLM agents through a bespoke sampling method. Subsequently, the generated data undergo a rigorous evaluation using a dual-model method that assesses both difficulty and quality. Finaly, the above process evolves in a dynamic refinement phase, where more effective LLMs are prioritized, enhancing the overall data quality. Our empirical studies, including instruction tuning experiments with models such as Pythia and LLaMA, demonstrate the effectiveness of the proposed framework. Optimized datasets have achieved substantial improvements, with an average increase of 12% and notable gains in specific metrics, such as a 40% improvement in Fermi, as evidenced by benchmarks like MT-bench, Vicuna bench, and WizardLM testset.

Figures (5)

• Figure 1: The diagram of the Star-Agents Framework. Step 1 is designed to gather diverse instructions and responses as shown in Appendix \ref{['subsec:data example']}. Step 2 focuses on selecting high-quality, tailored data from the data collected in Step 1. Finally, Step 3 aims to enhance the effectiveness and efficiency of the data generation process by evolving the Star-Agents framework.
• Figure 2: Performance comparison of varied-scale models on the Alpaca and Evol-Instruct datasets. The tasks from the Evol-Instruct dataset are more complex than those from Alpaca.
• Figure 3: Illustration of dual-model evaluation. Data with a significant gap between the IFD scores of the small and large models will be prioritised.
• Figure 4: Radar plot of detailed scores for Llama-2-7B-star_instrcut against the major baseline on different subtasks of (a) Vicuna-Bench and (b) MT-Bench.
• Figure 5: Evolution of the typical Agent-Pairs.