CoachLM: Automatic Instruction Revisions Improve the Data Quality in LLM Instruction Tuning

TL;DR

CoachLM is a novel method that automatically revises instruction datasets to significantly improve their quality, leading to better instruction-following performance in LLMs and more efficient data cleaning processes.

Contribution

It introduces CoachLM, a new approach for dataset enhancement through automatic revisions, increasing high-quality samples from 17.7% to 78.9% and boosting LLM performance.

Findings

High-quality instruction samples increased to 78.9%.

Instruction-following capabilities improved by 29.9%.

Data cleaning efficiency improved by 20%.

Abstract

Instruction tuning is crucial for enabling Language Learning Models (LLMs) in responding to human instructions. The quality of instruction pairs used for tuning greatly affects the performance of LLMs. However, the manual creation of high-quality instruction datasets is costly, leading to the adoption of automatic generation of instruction pairs by LLMs as a popular alternative. To ensure the high quality of LLM-generated instruction datasets, several approaches have been proposed. Nevertheless, existing methods either compromise dataset integrity by filtering a large proportion of samples, or are unsuitable for industrial applications. In this paper, instead of discarding low-quality samples, we propose CoachLM, a novel approach to enhance the quality of instruction datasets through automatic revisions on samples in the dataset. CoachLM is trained from the samples revised by human experts and significantly increases the proportion of high-quality samples in the dataset from 17.7% to 78.9%. The effectiveness of CoachLM is further assessed on various real-world instruction test sets. The results show that CoachLM improves the instruction-following capabilities of the instruction-tuned LLM by an average of 29.9%, which even surpasses larger LLMs with nearly twice the number of parameters. Furthermore, CoachLM is successfully deployed in a data management system for LLMs at Huawei, resulting in an efficiency improvement of up to 20% in the cleaning of 40k real-world instruction pairs. We release various assets of CoachLM, including the training data, code and test set (https://github.com/lunyiliu/CoachLM).