PEFT-Factory: Unified Parameter-Efficient Fine-Tuning of Autoregressive Large Language Models

TL;DR

PEFT-Factory presents a modular, open-source framework to unify and accelerate parameter-efficient fine-tuning of autoregressive LLMs. By supporting both off-the-shelf and custom PEFT methods, integrating 27 datasets across classification and generation tasks, and introducing the PSCP efficiency metric, it enables fair, reproducible benchmarking. The framework's plugin-style dynamic loading and dataset enhancements promote rapid prototyping and standardized evaluation, evidenced by use cases on reproducibility and benchmarking (e.g., the PEFT-Bench). Built atop LLaMA-Factory and leveraging PyTorch, PEFT-Factory aims to become a go-to platform for fair comparison and extensibility of PEFT methods in the NLP community.

Abstract

Parameter-Efficient Fine-Tuning (PEFT) methods address the increasing size of Large Language Models (LLMs). Currently, many newly introduced PEFT methods are challenging to replicate, deploy, or compare with one another. To address this, we introduce PEFT-Factory, a unified framework for efficient fine-tuning LLMs using both off-the-shelf and custom PEFT methods. While its modular design supports extensibility, it natively provides a representative set of 19 PEFT methods, 27 classification and text generation datasets addressing 12 tasks, and both standard and PEFT-specific evaluation metrics. As a result, PEFT-Factory provides a ready-to-use, controlled, and stable environment, improving replicability and benchmarking of PEFT methods. PEFT-Factory is a downstream framework that originates from the popular LLaMA-Factory, and is publicly available at https://github.com/kinit-sk/PEFT-Factory

Figures (5)

• Figure 1: Diagram representing the components of PEFT-Factory. The four main overarching components of PEFT-Factory are PEFT Methods, Datasets, Models, and Metrics, which are further defined by their subcomponents. Components represented by green color are implemented in PEFT-Factory, components in blue color are native to LLaMA-Factory zheng-etal-2024-llamafactory. Additionally, the Adapters library requires a different model class than the rest of the PEFT provider frameworks.
• Figure 2: Selection of PEFT methods from Finetuning method dropdown menu. All 19 PEFT methods included in PEFT-Factory are available to choose.
• Figure 3: Configuration options for the Prompt Tuning method.
• Figure 4: Classification and PSCP results for prediction after training with Prompt Tuning.
• Figure 5: Example directory structure of custom PEFT interface used for dynamic loading of PEFT custom methods.