RARR: Researching and Revising What Language Models Say, Using Language Models

Welcome! This repository contains a public implementation for RARR: Researching and Revising What Language Models Say, Using Language Models from CMU, Google Research, and UC Irvine. RARR addresses the attribution and factuality problems in large language models (LLMs) in a post-hoc manner by retrieving evidence to attribute the LLM's output and editing the output to be facutally consistent with the evidence. RARR uses large language models themselves to do the editing.

The techniques in RARR are quite general so if you are working on any task using LLMs to edit the output of LLMs (e.g., removing bias, fixing toxing outputs, etc.), you may find this repository a good starting point.

The easiest way to start using RARR is to NOT pull this repository, and instead copy this Colab notebook into your Google Drive and run the cells to get started! Alternatively, you can pull this repository and follow the following steps to get started.

Getting Started

The repository was tested on Ubuntu 20.04.2 LTS using Python 3.8. First install dependencies in requirements.txt, then run python -m spacy download en_core_web_sm.

Setting up APIs

Bing API

We use Bing to search for relevant evidence. To setup a Bing API, create an Azure account, then create a subscription, then create a Bing Search API using the subscription you created. We recommend the S2 pricing plan which costs $3 every 1000 calls. You will be given two API keys. Copy one of the keys and add the following to your ~/.bash_profile:

export AZURE_SEARCH_KEY="<BING_API_KEY>"

If you would like to quickly test the RARR pipeline without having access to a search API key, see [Editing a Single Claim](### Editing a Single Claim).

OpenAI API

We use LLMs from OpenAI to verify and edit the claims. Add the following to your ~/.bash_profile:

export OPENAI_API_KEY="<OPENAI_API_KEY>"

Using RARR to Improve Attribution and Factuality of LLMs

RARR (Retrofit Attribution using Research and Revision), improves the attribution and factuality of langauge models by taking their outputs and applying a post-hoc retrieve-and-edit approach. Given the output of a LLM (i.e., claim), RARR applies the following steps:

1. Question Generation: We generate a set of queries using a prompted LLM to interrogate the claim.
2. Search: For each query, we search for relevant webpages, then apply a passage extractor to retrieve the most relevant passage(s) for the query as evidence.
3. We iteratively use each evidence to edit the claim in two steps.
   1. Agreement Gate: We use a prompted LLM to decide whether the current evidence contradicts the information in the claim. If so, we move to editing. If not, we skip editing.
   2. Editing: We use a prompted LLM to edit the claim so it is consistent with the current evidence.
4. Generating Attribution Report: We extract a subset of the evidence as an attribution report (i.e., a citation for information in the claim).

Editing a File of Claims

Given a JSONLines file, where each line is a dictionary with the key input_info and under this is a dictionary with a claim field run the following. An example line: {"input_info": {"claim": "Michael Jordan played for the LA Lakers."}}

python run_editor_sequential.py \
  --input_file "path/to/input_file.jsonl" \
  --output_file "path/to/output_file.jsonl" \
  --model_name "text-davinci-003" \
  --claim_field "claim"

WARNING!! We also provide the ability to provide a --hallucinate-evidence flag which uses a LLM to generate evidence instead of retrieving it. We provide this flag to quickly test the repository quickly in the event a search API cannot be obtained. This flag should NEVER be set when using RARR to improve attribution as the evidence generated may contain hallucinations themselves.

Editing a Single Claim

import json
from run_editor_sequential import run_editor_one_instance

claim = "Michael Jordan played for the LA Lakers."
result = run_editor_one_instance(claim=claim, model="text-davinci-003")
print(json.dumps(result, indent=4))

# To hallucinate evidence using a LLM. Do NOT trust attribution results in this mode.
claim = "Michael Jordan played for the LA Lakers."
do_not_trust_result = run_editor_one_instance(claim=claim, model="text-davinci-003", hallucinate_evidence=True)
print(json.dumps(do_not_trust_result, indent=4))

Citation

If you find this repository useful, please cite the RARR paper.

@article{Gao2022RARRRA,
  title={RARR: Researching and Revising What Language Models Say, Using Language Models},
  author={Luyu Gao and Zhuyun Dai and Panupong Pasupat and Anthony Chen and Arun Tejasvi Chaganty and Yicheng Fan and Vincent Zhao and N. Lao and Hongrae Lee and Da-Cheng Juan and Kelvin Guu},
  journal={ArXiv},
  year={2022},
  volume={abs/2210.08726},
  url={https://arxiv.org/abs/2210.08726},
}