Deep Learning Library Extension
2.8k 2026-04-30
adapter-hub/adapters
A unified library extending HuggingFace Transformers for parameter-efficient and modular transfer learning in NLP.
Core Features
Integrates 10+ adapter methods into 20+ state-of-the-art Transformer models.
Supports full-precision or quantized training (e.g., Q-LoRA, Q-Bottleneck Adapters).
Enables adapter merging via task arithmetics and composition blocks.
Provides a unified interface for efficient fine-tuning and modular transfer learning.
Compatible with previously trained adapters from `adapter-transformers`.
Quick Start
pip install -U adaptersDetailed Introduction
The Adapters project is a powerful add-on library to HuggingFace's Transformers, designed to streamline parameter-efficient and modular transfer learning, particularly for NLP tasks. It integrates a wide array of adapter methods into numerous state-of-the-art Transformer models, offering a unified interface for efficient fine-tuning. The library supports advanced features like quantized training, adapter merging, and complex adapter composition, enabling researchers and developers to explore cutting-edge techniques in deep learning with minimal coding overhead.