Long-Form Factuality in Large Language Models

This is the official code release accompanying our paper "Long-form factuality in large language models". This repository contains:

1. LongFact: A prompt set of 2,280 fact-seeking prompts requiring long-form responses.
2. Search-Augmented Factuality Evaluator (SAFE): Automatic evaluation of model responses in long-form factuality settings.
3. F1@K: Extending F1 score to long-form settings using recall from human-preferred length.
4. Experimentation pipeline for benchmarking OpenAI and Anthropic models using LongFact + SAFE.

Installation

First, clone our GitHub repository.

git clone https://github.com/google-deepmind/long-form-factuality.git

Then navigate to the newly-created folder.

cd long-form-factuality

Next, create a new Python 3.10+ environment using conda.

conda create --name longfact python=3.10

Activate the newly-created environment.

conda activate longfact

All external package requirements are listed in requirements.txt. To install all packages, and run the following command.

pip install -r requirements.txt

Usage

LongFact

The full prompt set for LongFact is available in the longfact/ folder. See the README in longfact/ for more details about the dataset.

To run the data-generation pipeline that we used to generate LongFact, use the following command. Refer to the README in data_creation/ for additional details about the data-generation pipeline.

python -m data_creation.pipeline

SAFE

Our full implementation of SAFE is located in eval/safe/. See the README in eval/safe/ for more information about how SAFE works.

To run the pipeline for evaluating SAFE against FActScore human annotations, use the following command. Refer to the README in eval/ for additional details about this experiment.

python -m eval.correlation_vs_factscore

Benchmarking models

To benchmark OpenAI and Anthropic models, first add your API keys to common/shared_config.py (see README in common/ for more information; be sure not to publish these keys). To obtain model responses for a given prompt set, use the following command. Refer to the README in main/ for additional details about our main experimentation pipeline.

python -m main.pipeline

Next, to evaluate prompt-response pairs from our main experimentation pipeline using SAFE, use the following command, making sure to add the path to the .json file containing the prompt-response pairs to be evaluated to the --result_path argument.

python -m eval.run_eval \
    --result_path=

Unit Tests

Each file in this directory has a corresponding unit test with the _test suffix (e.g., file.py would have file_test.py for unit tests). Run commands for individual tests are shown in the unit test files. To run all unit tests, use the following command.

python -m unittest discover -s ./ -p "*_test.py"

Citing this work

If you find our code useful, please cite our paper:

@article{wei2024long,
  title={Long-form factuality in large language models},
  author={Wei, Jerry and Yang, Chengrun and Song, Xinying and Lu, Yifeng and Hu, Nathan and Huang, Jie and Tran, Dustin and Peng, Daiyi and Liu, Ruibo and Huang, Da and Du, Cosmo and Le, Quoc V},
  journal={arXiv preprint arXiv:2403.18802},
  year={2024}
}

License and disclaimer

Copyright 2024 DeepMind Technologies Limited

All software is licensed under the Apache License, Version 2.0 (Apache 2.0); you may not use this file except in compliance with the Apache 2.0 license. You may obtain a copy of the Apache 2.0 license at: https://www.apache.org/licenses/LICENSE-2.0

All other materials are licensed under the Creative Commons Attribution 4.0 International License (CC-BY). You may obtain a copy of the CC-BY license at: https://creativecommons.org/licenses/by/4.0/legalcode

Unless required by applicable law or agreed to in writing, all software and materials distributed here under the Apache 2.0 or CC-BY licenses are distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the licenses for the specific language governing permissions and limitations under those licenses.

This is not an official Google product.