adapters

Why we included this project

Fine-tuning a Transformer model usually means updating every parameter, which gets expensive fast. Adapters takes a different route by adding small trainable modules to a frozen base model, so you only train a fraction of the weights. The library sits on top of HuggingFace Transformers and bundles more than ten parameter-efficient methods, from bottleneck adapters and LoRA to prefix tuning, behind one consistent API. Switching between them is a configuration change rather than a rewrite of your training loop. Because each adapter is compact, you can train several for different tasks against one shared base model and store or deploy them separately. It also supports quantized training like Q-LoRA, adapter merging via task arithmetic, and composing multiple adapters, which is useful for serious work on modular transfer learning.