adapter-hub/adapters

Health v2 · maintenance only

Adoption not part of the score

2826 stars · 373 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-29, confidence not recorded

Adapters is a Python add-on library for HuggingFace Transformers that integrates 10+ parameter-efficient fine-tuning methods (bottleneck adapters, LoRA, Prefix Tuning, DoRA, etc.) into 20+ Transformer models. It provides a unified interface for training, composing, merging, and sharing adapter modules for modular transfer learning in NLP.

Use cases

• fine-tune large language models with LoRA or QLoRA on limited GPU memory
• train multiple task adapters on one shared base Transformer model
• compose or merge adapters via task arithmetics and composition blocks
• run quantized parameter-efficient training like Q-LoRA on Llama models
• share and load pre-trained adapter modules from a central hub
• research parameter-efficient transfer learning methods for NLP
• swap task-specific heads and adapters dynamically for multi-task inference

When to choose

• you already use HuggingFace Transformers and want adapters or LoRA-style PEFT with a unified API
• you need to train, stack, fuse, or merge multiple adapters on shared base models
• you want quantized low-memory fine-tuning of large language models
• you need compatibility with adapters trained via the legacy adapter-transformers package

When to avoid

• you need full fine-tuning of all model parameters rather than parameter-efficient methods
• your work is outside NLP/Transformer language models
• you prefer a different PEFT ecosystem such as HuggingFace PEFT and don't need adapter composition features

Facets

library · maturity active

machine-learning llm-training nlp machine-learning deep-learning large-language-models python parameter-efficient-fine-tuning peft adapters lora transformers huggingface transfer-learning pytorch natural-language-processing

Member repositories

For agents

markdown · JSON · MCP: product_card(name="adapter-hub/adapters")