LIMIT: Less Is More for Instruction Tuning Across Evaluation Paradigms

TL;DR

This paper demonstrates that small, carefully selected instruction datasets can effectively fine-tune large language models for diverse evaluation tasks, challenging the notion that large datasets are necessary.

Contribution

It shows that limited, high-quality instruction data can achieve competitive performance across multiple evaluation paradigms, offering a more efficient finetuning approach.

Findings

Small datasets of 1k-6k samples suffice for good performance.

Mixing different dataset types improves overall results.

Efficient finetuning is possible without large datasets.

Abstract

Large Language Models are traditionally finetuned on large instruction datasets. However recent studies suggest that small, high-quality datasets can suffice for general purpose instruction following. This lack of consensus surrounding finetuning best practices is in part due to rapidly diverging approaches to LLM evaluation. In this study, we ask whether a small amount of diverse finetuning samples can improve performance on both traditional perplexity-based NLP benchmarks, and on open-ended, model-based evaluation. We finetune open-source MPT-7B and MPT-30B models on instruction finetuning datasets of various sizes ranging from 1k to 60k samples. We find that subsets of 1k-6k instruction finetuning samples are sufficient to achieve good performance on both (1) traditional NLP benchmarks and (2) model-based evaluation. Finally, we show that mixing textbook-style and open-ended QA finetuning datasets optimizes performance on both evaluation paradigms.