Lorahub
Project description
LoraHub: Efficient Cross-Task Generalization via Dynamic LoRA Composition
The official repository which contains the code and pre-trained models for our paper LoraHub: Efficient Cross-Task Generalization via Dynamic LoRA Composition.
🔥 Updates
- [2023-8-29]: We released the full produce code at reproduce_bbh.py. Please checkout the script to reproduce our results!
- [2023-8-03]: Integrated into Replicate, check out the demo!
- [2023-7-27]: We released our code and demo. Check it out!
- [2023-7-26]: We released our paper.
🏴 Overview
Low-rank adaptations (LoRA) are techniques for fine-tuning large language models on new tasks. We propose LoraHub, a framework that allows composing multiple LoRA modules trained on different tasks. The goal is to achieve good performance on unseen tasks using just a few examples, without needing extra parameters or training. And we want to build a marketplace where users can share their trained LoRA modules, thereby facilitating the application of these modules to new tasks.
The figure demostrates the zero-shot learning, few-shot in-context learning and few-shot lorahub learning (ours). Note that the Compose procedure is conducted per task rather than per example. Our method achieves similar inference throughput as zero-shot learning, yet approaches the performance of in-context learning on the BIG-Bench Hard (BBH) benchmark. The experimental results show the superior efficacy of our method in comparison to zero-shot learning while closely resembling the performance of in-context learning (ICL) in few-shot scenarios.
The figure shows the pipeline of LoraHub Learning. Our method encompasses two stages: the Compose stage and the Adapt stage. During the Compose stage, existing LoRA modules are integrated into one unified module, employing a set of weights, denoted as w, as coefficients. In the Adapt stage, the amalgamated LoRA module is evaluated on a few examples from the unseen task. Subsequently, a gradient-free algorithm is applied to refine w. After executing K iterations, a highly adapted LoRA module is produced, which can be incorporated with the LLM to perform the intended task.
🌲 Project Structure
Our code is organized as below:
|-- lorahub
-- algorithm.py # main code for lorahub learning and inference
-- constant.py # lora candidate module names
|-- example.py # usage code for demonstration purpose
And you can use LoraHub learning by simply calling the following function:
from lorahub.algorithm import lorahub_learning
lorahub_learning(lora_module_list: List[str], # list of lora candidates
example_inputs: List[str],
example_outputs: List[str],
max_inference_step: int,
model_name_or_path=None, # if not given, we will use the model_name_or_path in lora config
batch_size=None,
get_loss=default_get_loss, # The function to get the objective for optimiztion, use loss as default (can be changed to something like acc. or similarity)
get_regular=default_l1_regularization, # The function to get regularization term for the weight, use 0.05*|w_i| as default
seed=42)
⚡️ Quickstart
Prepare Environment
First, you should run the following commands to install the latest lib developed for LoraHub.
pip install datasets
pip install transformers
pip install peft
pip install nevergrad
pip install torch
pip install tqdm
pip install pandas
pip install numpy
Install LoraHub
The pypi package will be released in a few days. Please stay tuned!
🏰 Resource
LoRA Candidates
Our methodology requires a compendium of LoRA modules trained on preceding tasks. For parity with Flan, we adopt the tasks utilized to instruct Flan-T5, thereby incorporating nearly 196 distinct tasks and their corresponding instructions via https://huggingface.co/datasets/conceptofmind/FLAN_2022. Following this, we created several LoRA modules as possible candidates. These LoRA modules can be accessed at https://huggingface.co/models?search=lorahub.
💬 Citation
If our work is useful for you, please consider citing our paper:
@misc{huang2023lorahub,
title={LoraHub: Efficient Cross-Task Generalization via Dynamic LoRA Composition},
author={Chengsong Huang and Qian Liu and Bill Yuchen Lin and Tianyu Pang and Chao Du and Min Lin},
year={2023},
eprint={2307.13269},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
File details
Details for the file lorahub-0.1.0.tar.gz
.
File metadata
- Download URL: lorahub-0.1.0.tar.gz
- Upload date:
- Size: 12.0 kB
- Tags: Source
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/4.0.2 CPython/3.9.13
File hashes
Algorithm | Hash digest | |
---|---|---|
SHA256 | 83cf9eac4a07344e960bfedc3812b6e3198259d9832e38b1f5e239ca6b02a660 |
|
MD5 | 83230f9a8be59fe14d4e24243727fe9b |
|
BLAKE2b-256 | 74bc4506162792b64b14c4b35cd12e5d7d8348f3e88f69afdad3ab8efd45e0be |
File details
Details for the file lorahub-0.1.0-py3-none-any.whl
.
File metadata
- Download URL: lorahub-0.1.0-py3-none-any.whl
- Upload date:
- Size: 9.9 kB
- Tags: Python 3
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/4.0.2 CPython/3.9.13
File hashes
Algorithm | Hash digest | |
---|---|---|
SHA256 | d98405c0b72d61f14faf111e9f71313372cbea72a34f9942766741d39898d854 |
|
MD5 | c2e0bdde4c1a38cd7ff5d662e2f202ef |
|
BLAKE2b-256 | b98dc9748fa2f8d111f6f8457ced53a86568df5b76f5fb949604ce5d6c5d7ce7 |