Awesome
FIGA
This repository is the official implementation of ICLR 2024 paper: Beyond Imitation: Leveraging Fine-grained Quality Signals for Alignment.
Quick Start
Considering that a modified version of transformers will be installed, it is recommended to create a new conda environment:
conda create -n FIGA python=3.8
conda activate FIGA
conda install pytorch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 pytorch-cuda=12.1 -c pytorch -c nvidia
You should clone the FIGA repository and follow its instructions.
git clone https://github.com/RUCAIBox/FIGA.git && cd FIGA
pip install -r requirements.txt
After this, you need to replace the trainer_utils.py
and modeling_llama.py
files in the transformers library with the corresponding files from this repository. This is necessary for fine-tuning using the FIGA method.
SPA Dataset
You can download SPA dataset in: https://huggingface.co/datasets/RUCAIBox/SPA.
For our publicly available SPA dataset, the output
field is the ground truth response, the original_output
field contains results generated by the alpaca-7b model, and the revised_output
field contains results modified by using a more powerful model (i.e. ChatGPT-3.5). For a detailed description of the construction process of the SPA dataset, please refer to our paper.
Instruction tuning
After setting up the environment, you can utilize the FIGA method to fine-tune the model by referring to the code provided below:
bash bash/run_7b.sh > output.log 2>&1
Acknowledgment
Please cite the following paper if you find our code or data helpful.
@article{guo2023beyond,
title={Beyond imitation: Leveraging fine-grained quality signals for alignment},
author={Guo, Geyang and Zhao, Ranchi and Tang, Tianyi and Zhao, Wayne Xin and Wen, Ji-Rong},
journal={arXiv preprint arXiv:2311.04072},
year={2023}
}