2.5 KiB
🦙🌲🤏 Alpaca-LoRA: Low-Rank LLaMA Instruct-Tuning
This repository contains code for reproducing the Stanford Alpaca results using low-rank adaptation (LoRA).
The fine-tuning runs within five hours on a consumer GPU,
and the LoRA weights are made available on the Huggingface model hub.
With Huggingface's out-of-the-box 8-bit quantization,
we aim to provide an Instruct model of similar quality to text-davinci-003 that can run on a Raspberry Pi. (For research.)
Until Jason Phang's LLaMA implementation
is merged, users will need to replace their local Huggingface transformers as described below.
For fine-tuning LoRAs we use Huggingface's PEFT.
Included also is code to download the LLaMA foundation model from the Huggingface model hub. (For research.)
Once I've finished running the finetuning code myself, I'll put the LoRA on the Hub as well, and the code in generate.py should work as expected.
Setup
- Install dependencies (install zphang's transformers fork)
pip install -q datasets loralib sentencepiece
pip uninstall transformers
pip install -q git+https://github.com/zphang/transformers@c3dc391
pip install -q git+https://github.com/huggingface/peft.git
Inference (generate.py)
See generate.py. This file reads the decapoda-research/llama-7b-hf model from the Huggingface model hub and the LoRA weights from tloen/alpaca-lora-7b, and runs inference on a specified input. Users should treat this as example code for the use of the model, and modify it as needed.
Training (finetune.py)
Under construction. If you're impatient, note that this file contains a set of hardcoded hyperparameters you should feel free to modify. PRs adapting this code to multi-GPU setups and larger models are always welcome.
To do
- Merge LoRA weights into LLaMA weights to remove inference dependency on PEFT
- Train/val/test split
- Hyperparameter tuning code
- Documentation for notebook
- Support for
13b,30b,65b - Train a version that doesn't waste tokens on the prompt header
- Inference CLI and evaluation
- Better disclaimers about why using LLaMA without permission is very bad!