Demystifying Instruction Mixing for Fine-tuning Large Language Models

Renxi Wang, Haonan Li, Minghao Wu, Yuxia Wang, Xudong Han +2 more
12/17/2023
cs.CLcs.AI

No official code found. Best implementation: Reason-Wang/InstructLLM (4 stars, last commit 2024-01-05).

Abstract

Instruction tuning significantly enhances the performance of large language models (LLMs) across various tasks. However, the procedure to optimizing the mixing of instruction datasets for LLM fine-tuning is still poorly understood. This study categorizes instructions into three primary types: NLP downstream tasks, coding, and general chat. We explore the effects of instruction tuning on different combinations of datasets on LLM performance, and find that certain instruction types are more advantageous for specific applications but can negatively impact other areas. This work provides insights into instruction mixtures, laying the foundations for future research.

View on arXivView PDF

Code Implementations(1)

Reason-Wang/InstructLLMCommunity100%

The official implementation of paper "Demystifying Instruction Mixing for Fine-tuning Large Language Models"

40Shell, PythonJan 4, 20242 years ago
deepspeedfine-tuninginstruction-tuningllama2llm+2 more

Similar Papers

From Base to Conversational: Japanese Instruction Dataset and Tuning Large Language Models

93% match1 repo72023

Maybe Only 0.5% Data is Needed: A Preliminary Exploration of Low Training Data Instruction Tuning

92% match2023

MLLM-Selector: Necessity and Diversity-driven High-Value Data Selection for Enhanced Visual Instruction Tuning

92% match2025

Instruction Tuning for Large Language Models: A Survey

91% match1 repo2322023

Dynamics of Instruction Fine-Tuning for Chinese Large Language Models

90% match1 repo182023

An Empirical Study of Instruction-tuning Large Language Models in Chinese

90% match1 repo2.8k2023

Cite this paper

@article{wang2023demystifying,
  title  = {Demystifying Instruction Mixing for Fine-tuning Large Language Models},
  author = {Renxi Wang and Haonan Li and Minghao Wu and Yuxia Wang and Xudong Han and Chiyu Zhang and Timothy Baldwin},
  year   = {2023},
  eprint = {2312.10793},
  archivePrefix = {arXiv},
  url    = {http://arxiv.org/abs/2312.10793v3}
}

Add a badge to your README

Shows whether official code is linked for this paper.

Code of Paper badge
[![Code of Paper](https://codeofpaper.com/badge/2312.10793.svg)](https://codeofpaper.com/paper/2312.10793)

Discussion