RA-DIT: Retrieval-Augmented Dual Instruction Tuning

This is the official artifact release of the follwoing paper:

Xi Victoria Lin*, Xilun Chen*, Mingda Chen*, Weijia Shi, Maria Lomeli, Rich James, Pedro Rodriguez, Jacob Kahn, Gergely Szilvasy, Mike Lewis, Luke Zettlemoyer, Wen-tau Yih. RA-DIT: Retrieval-Augmented Dual Instruction Tuning, ICLR 2024.

Overview

We are releasing the following artifacts from the paper.

1. The scripts for processing the fine-tuning datasets into the format used in the paper.
2. The DRAGON (Lin et al. 2023) embeddings of the Wiki 2021 index† used for fine-tuning datasets generation and evaluation.

†For the benchmark datasets used in our paper, using the Wiki 2021 index retains the majority of performance benefits, comparable to using the full retrieval index that combines both Wikipedia and Common Crawl data (Appendix E).

Conda Setup

We recommend cloning the Atlas codebase (Izcard et al. 2022) and starting from its conda building instructions outlined here, since we will be using its retriever implementation.

If you plan to use your own retriever and just need to run data generation, you may keep using an existing conda or create a vanilla Python environment:

conda create --name radit-env python=3.8
conda activate radit-env

Lastly, install RA-DIT dependencies in your conda

pip install -r requirements.txt

Retriever Setup (WIP)

We used the distributed index implementation by Izacard et al. 2022 for passage retrieval†.

†You can serve the passages and embeddings using any efficient vector search library, such as faiss.

git clone https://github.com/facebookresearch/atlas.git

Dataset Preparation

We recommend creating a root directory on your disk to save all the intermediate and final processed datasets, e.g.

export DATA_ROOT=/tmp/radit_datasets/
mkdir $DATA_ROOT

Step 1: Download Hugging Face Datasets

All our training and evaluation task datasets can be downloaded from Hugging Face. To download the dataset,

python3 src/data/download_datasets.py --root_dump_dir $DATA_ROOT/raw

Step 2: Prepare Language Model Fine-tuning Data

1. Run offline retrieval

PYTHONPATH=. python3 src/data/lm_ft/prepare_lm_ft_datasets.py --input_dir $DATA_ROOT/raw --save_dir $DATA_ROOT/raw.ra --action retrieval

2. Format and save retrieval-augmented (RA) instruction-tuning (IT) data

PYTHONPATH=. python3 src/data/lm_ft/prepare_lm_ft_datasets.py --input_dir $DATA_ROOT/raw --save_dir $DATA_ROOT/ft.ra --n_support 3 --action format_data

3. Format and save instruction-tuning (IT) data with no retrieval

PYTHONPATH=. python3 src/data/lm_ft/prepare_lm_ft_datasets.py --input_dir $DATA_ROOT/raw --save_dir $DATA_ROOT/ft --action format_data --ignore_retrieved_passages

Step 3: Prepare Retriever Fine-tuning Data

PYTHONPATH=. python3 src/data/retriever_ft/prepare_retriever_ft_datasets.py --input-dir $DATA_ROOT/raw --save-dir  $DATA_ROOT/retriever_ft/lsr_score_saving_input/ --task-list retriever_train

Citation

If you find the resource in this repository helpful, please cite

@inproceedings{DBLP:conf/iclr/Lin0CSL00KSLZY24,
  author       = {Xi Victoria Lin and
                  Xilun Chen and
                  Mingda Chen and
                  Weijia Shi and
                  Maria Lomeli and
                  Richard James and
                  Pedro Rodriguez and
                  Jacob Kahn and
                  Gergely Szilvasy and
                  Mike Lewis and
                  Luke Zettlemoyer and
                  Wen{-}tau Yih},
  title        = {{RA-DIT:} Retrieval-Augmented Dual Instruction Tuning},
  booktitle    = {The Twelfth International Conference on Learning Representations,
                  {ICLR} 2024, Vienna, Austria, May 7-11, 2024},
  publisher    = {OpenReview.net},
  year         = {2024},
  url          = {https://openreview.net/forum?id=22OTbutug9},
  timestamp    = {Wed, 07 Aug 2024 17:11:53 +0200},
  biburl       = {https://dblp.org/rec/conf/iclr/Lin0CSL00KSLZY24.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}

License

The code in this repository is licensed under CC-BY-NC.