Becoming Self-Instruct: Introducing Early Stopping Criteria for Minimal Instruct Tuning

By Waseem AlShikh et al

Read the original document by opening this link in a new tab.

Table of Contents

Abstract
1 Introduction
2 Background and Related Work
3 Instruction Following Index
4 Binary classifier and Instruction Following Score
5 Results
5.1 Baseline
5.2 Prompt Engineering
5.3 Supervised Finetuning

Summary

In this paper, the authors introduce the Instruction Following Score (IFS) metric to detect language models' ability to follow instructions. The paper discusses the distinction between base and instruct models, proposing IFS as an early stopping criterion for instruct tuning. The study evaluates publicly available models with IFS scores and explores prompt engineering techniques to improve instruction following. Additionally, supervised finetuning experiments on 7B and 13B LLaMA models are conducted using the gpt4all v1.3-groovy dataset. The results show significant improvements in instruction following scores with prompt suffixes and supervised finetuning, indicating the effectiveness of these methods in encouraging models to follow instructions.