lorahub

Why we included this project

Most teams fine-tune a fresh LoRA adapter for each new task, which gets expensive as the number of tasks grows. LoraHub takes a different route: it pulls together LoRA modules already trained on other tasks and blends them into a single adapter, guided by just a handful of examples from the target task. The compose-then-adapt pipeline searches over module weights without any gradient updates, so you adapt to a new task without adding trainable parameters. The repo is a clean research implementation, pip-installable, with a small API, the full reproduction script, and the Flan-based LoRA candidates used in the paper. It's a useful reference if you're exploring how to reuse and share adapters across tasks, though it's a research artifact rather than a maintained production service.