LoraHub is a framework designed to enhance the efficiency of fine-tuning large language models (LLMs) for new tasks by dynamically composing pre-trained Low-Rank Adaptation (LoRA) modules. It targets researchers and developers seeking to achieve strong few-shot performance on unseen tasks without requiring additional training or parameters, effectively acting as a "marketplace" for reusable LoRA modules.

How It Works

LoraHub employs a two-stage process: Compose and Adapt. In the Compose stage, multiple LoRA modules trained on different tasks are integrated into a single, unified module using learned coefficients. The Adapt stage then refines these coefficients on a few examples from the target task using a gradient-free optimization algorithm. This approach aims to match the performance of in-context learning (ICL) while maintaining the inference throughput of zero-shot learning.

Quick Start & Requirements

• Install via pip: pip install lorahub
• Requires a collection of pre-trained LoRA modules. The project suggests using modules trained on Flan-T5 tasks, available at Hugging Face (https://huggingface.co/models?search=lorahub).
• Example usage is provided in example.py.

Highlighted Details

• Achieves performance close to in-context learning (ICL) on the BIG-Bench Hard (BBH) benchmark with similar inference throughput to zero-shot learning.
• Enables cross-task generalization by composing existing LoRA modules.
• Uses a gradient-free algorithm to adapt composed LoRA modules to new tasks.
• Supports customization of the loss function and regularization terms.

Maintenance & Community

• The project paper was accepted by COLM 2024.
• Code and demo are available.
• Integrated into Replicate.

Licensing & Compatibility

• The README does not explicitly state the license.

Limitations & Caveats

• The effectiveness relies on the availability and quality of pre-trained LoRA modules.
• The project is presented as a research artifact, and extensive production-readiness is not detailed.