pyreft

Why we included this project

pyreft comes from the Stanford NLP group and implements Representation Fine-Tuning (ReFT), a lighter alternative to LoRA and adapter-style methods. Where those approaches add trainable weights across every layer and timestep, ReFT learns small interventions on the model's hidden representations, and only at selected positions. That narrower touch keeps memory and compute costs down, and the library works with any pretrained model you already have on HuggingFace. You get training support, hyperparameter configs, and the option to push results back to the Hub, so it drops into a fairly ordinary fine-tuning workflow. If you have been sizing LoRA ranks and wonder whether intervening on representations can match that accuracy with even less trainable surface, it is worth a quick experiment in a notebook.