amazon-science/synthesizrr

amazon-science/synthesizrr

Description: Synthesizing realistic and diverse text-datasets from augmented LLMs

Language: Python

License: Apache-2.0

Stars: 19

Forks: 5

Open issues: 8

Created: 2024-06-20T15:13:16Z

Pushed: 2026-04-04T08:57:03Z

Default branch: main

Fork: no

Archived: no

README:

SynthesizRR: Generating Diverse Datasets with Retrieval Augmentation

This repository contains the implementation of the paper "SynthesizRR: Generating Diverse Datasets with Retrieval Augmentation" (https://arxiv.org/abs/2405.10040)

Our proposed approach is below. Refer Algorithm 1 in the paper for details: https://arxiv.org/abs/2405.10040


Installing dependencies

We recommend installing required dependencies in a new Conda environment using the commands below.

These commands were tested to work on Deep Learning AMI GPU PyTorch 1.13.1 (Amazon Linux 2) 20230221 from AWS.

Install dependencies:

conda create -n synthesizrr python=3.11.8 --yes
conda activate synthesizrr
pip install uv ## For super-fast installation

uv pip install -r requirements.txt

uv pip install "spacy==3.7.4" "spacy-transformers==1.3.5"
uv pip install "setuptools==69.5.1"

python -m spacy download en_core_web_lg
python -c "import nltk; nltk.download('punkt');"

Code structure

synthesizrr/base/ contains utility functions and classes.

synthesizrr/expts/ contains code to reproduce the experiments.

Running the code

1. Setup DATA_DIR:

• Download the datasets into a local folder DATA_DIR.
• Inside synthesizrr/expt/data.py, set the variable DATA_DIR (marked TODO) to the above folder.

2. Setup CORPUS_DIR:

• Download the corpora into a folder CORPUS_DIR.
• We recommend using S3 for this since the corpora are large.
• Inside synthesizrr/expt/corpus.py, set the variable CORPUS_DIR (marked TODO) to the above folder.

3. Setup RESULTS_DIR:

• Inside synthesizrr/expt/common.py, set the variable RESULTS_DIR (marked with TODO) to a different folder. Intermediate datasets and metrics will be saved here.
• We recommend using S3 for this since the file-paths are long.

4. Start a Ray cluster:

• On the Ray head node, run: ray start --head
• On the Ray worker nodes, run ray start --address=':6379'
• At the top of the files data.py, corpus.py, main.py, add the following to connect to the Ray cluster:

import synthesizrr
import ray
from ray.util.dask import ray_dask_get, enable_dask_on_ray, disable_dask_on_ray
from pprint import pprint
pprint(ray.init(
address='ray://:10001', ## MODIFY THIS
ignore_reinit_error=True,
_temp_dir=str('/tmp/ray/'),
runtime_env={"py_modules": [
synthesizrr,
]},
))
enable_dask_on_ray()
pprint(ray.cluster_resources()) ## Shows you number of cpus and gpus to make sure it is setup properly.

5. After modifying the code to set DATA_DIR, CORPUS_DIR and RESULTS_DIR, and starting the Ray cluster, run the following:

• First, run cd synthesizrr/expts/ && python3 data.py to create the datasets. (You will need to download certain datasets to DATA_DIR folder beforehand).
• Next, run cd synthesizrr/expts/ && python3 corpus.py to create the corpora (warning, this step needs a lot of compute! Make sure you setup the Ray cluster and use a big machine with at least a few hundred GB of RAM as the head node).
• Finally, run the file cd synthesizrr/expts/ && python3 main.py to reproduce the experiments.

Security

See [CONTRIBUTING](CONTRIBUTING.md#security-issue-notifications) for more information.

License

This project is licensed under the Apache-2.0 License.

Citing

If you use or refer to this code in another publication, please cite it using the Bibtex below:

@misc{divekar2024synthesizrr,
title={SynthesizRR: Generating Diverse Datasets with Retrieval Augmentation},
author={Abhishek Divekar and Greg Durrett},
year={2024},
eprint={2405.10040},
archivePrefix={arXiv}
}

Acknowledgements

The compute infrastructure used for these experiments was financially supported by the Amazon Central Machine Learning department.

The following people contributed to the design or implemented smaller components in this codebase:

• Gaurav Manchanda
• Siba Rajendran
• Vijit Malik