# LLM-Adapters: An Adapter Family for Parameter-Efficient Fine-Tuning of Large Language Models

**Authors:** Zhiqiang Hu, Yihuai Lan, Lei Wang, Wanyu Xu, Ee-Peng Lim, Roy Ka-Wei Lee, Lidong Bing, Soujanya Poria  
**arXiv:** [2304.01933](https://arxiv.org/abs/2304.01933)  
**Date:** 2023-04-04  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.01933v1/

## Key points

- Natural Language Processing (NLP) has advanced with the development of large language models (LLMs) such as GPT-3, BLOOM, and LLaMA
- Fine-tuning is a popular technique used to adapt LLMs to specific downstream tasks by training them on task-specific datasets
- Closed-source nature of some powerful instruction-following LLMs limits their use for developing fine-tuning methods for specific downstream tasks
- Open access LLMs have been developed that can be fine-tuned using adapter based parameter efficient fine tuning (PEFT) methods
- Adapter based PEFT requires only fine tuning a few external parameters instead of the entire LLMs while achieving comparable or even better performance
- A framework called LLM Adapters has been presented that integrates various adapters into LLMs and can execute these adapter-based PEFT methods for different tasks
- The framework includes state-of-the-art open access LLMs such as LLaMA, BLOOM, OPT and GPT J, as well as widely used adapters such as Series adapter, Parallel adapter and LoRA.
- Experiments conducted on six datasets showed that using adapter-based PEFT in smaller scale LLMs yields comparable or even superior performance to that of powerful LLMs in zero-shot inference on simple math reasoning datasets.
- The proposed LLM Adapters framework provides a promising solution for fine-tuning large language models on downstream tasks.

## Summary

The field of Natural Language Processing (NLP) has seen significant advancements in recent years, particularly with the development of large language models (LLMs) such as GPT-3, BLOOM, and LLaMA. These models have demonstrated impressive performance on various NLP tasks. Fine-tuning is a popular technique used to adapt LLMs to specific downstream tasks by training them on task-specific datasets. However, the closed-source nature of some of the most powerful instruction-following LLMs like ChatGPT1 and GPT-42 limits their use as backbone models for developing fine-tuning methods for specific downstream tasks. To address this issue, researchers have developed open access LLMs that can be fine tuned with task specific data or instruction data using adapter based parameter efficient fine tuning (PEFT) methods. Adapter based PEFT requires only fine tuning a few external parameters instead of the entire LLMs while achieving comparable or even better performance. Among various fine tuning methods, adapter based PEFT is one of the most attractive topics. To enable further research on PEFT methods of LLMs, a team of researchers has presented an easy to use framework called LLM Adapters that integrates various adapters into LLMs and can execute these adapter based PEFT methods for different tasks. The framework includes state of the art open access LLMs such as LLaMA, BLOOM, OPT and GPT J, as well as widely used adapters such as Series adapter, Parallel adapter and LoRA. LLM Adapters is designed to be research friendly, efficient modular and extendable allowing integration of new adapters and evaluation with new larger scale LLMs. To evaluate the effectiveness of adapters in LLM Adapters for math reasoning datasets experiments were conducted on six datasets which showed that using adapter based PEFT in smaller scale LLMs (7B) with few extra trainable parameters yields comparable or even superior performance to that of powerful LLMs (175B) in zero shot inference on simple math reasoning datasets. Overall, the proposed LLM Adapters framework provides a promising solution for fine tuning large language models on downstream tasks. The framework is expected to advance adapter based PEFT research and facilitate the deployment of research pipelines while enabling practical applications to real world systems. The success and potential of this approach have sparked a range of adaptations and applications including Chinese Alpaca, Japanese Alpaca Thai Alpaca medical Alpaca (ChatDoctor), movie recommendation Alpaca (RecAlpaca), multi modal Alpaca(LLaMA Adapter), and LMFlow toolbox for fine tuning language models.