OpenFactCheck: Building, Benchmarking Customized Fact-Checking Systems and Evaluating the Factuality of Claims and LLMs

TL;DR

OpenFactCheck tackles the challenge of evaluating factuality in open-domain LLM outputs and comparing fact-checking approaches by introducing a unified, extensible framework with three modules: CustChecker for customizable checkers, LLMEval for standardized factuality evaluation, and CheckerEval for benchmarking verifiers. The framework supports a plugin-like, YAML-configured pipeline that can incorporate diverse claim processors, retrievers, and verifiers, and it is complemented by FactQA-based evaluation and a public leaderboard. Empirical results with LLaMA-2 and GPT-4 reveal that while GPT-4 generally yields high factuality, large portions of errors stem from snowballing, unknowns, and up-to-date information challenges, and that retrieval strategy and LLM prompts dominate performance and cost. The work provides a practical, open-source platform to standardize evaluation, facilitate fair comparisons, and accelerate progress in LLM factuality and automated fact-checking for researchers, developers, and search-enabled systems.

Abstract

The increased use of large language models (LLMs) across a variety of real-world applications calls for mechanisms to verify the factual accuracy of their outputs. Difficulties lie in assessing the factuality of free-form responses in open domains. Also, different papers use disparate evaluation benchmarks and measurements, which renders them hard to compare and hampers future progress. To mitigate these issues, we propose OpenFactCheck, a unified framework for building customized automatic fact-checking systems, benchmarking their accuracy, evaluating factuality of LLMs, and verifying claims in a document. OpenFactCheck consists of three modules: (i) CUSTCHECKER allows users to easily customize an automatic fact-checker and verify the factual correctness of documents and claims, (ii) LLMEVAL, a unified evaluation framework assesses LLM's factuality ability from various perspectives fairly, and (iii) CHECKEREVAL is an extensible solution for gauging the reliability of automatic fact-checkers' verification results using human-annotated datasets. Data and code are publicly available at https://github.com/yuxiaw/openfactcheck.

Figures (11)

• Figure 1: Overview of the OpenFactCheck framework with three modules. Green CustChecker: a customized fact-checker to identify factual errors given the outputs of LLMs. Orange LLMEval: a unified LLM factuality evaluator to assess the LLM factual ability from different aspects, and then to produce a report highlighting the domains and types of factual errors the model frequently makes, along with the improvement advice. Purple CheckerEval: a fact-checker evaluator and leaderboard to encourage the development of advanced checkers in terms of performance, time and costs.
• Figure 2: Automatic evaluation results for LLaMA-2 7B, 13B and GPT-4 responses on datasets of FacTool-QA, FELM-WK, and Factcheck-Bench using FacTool. left: the percentage of true claims, center: the number of false claims, and right: the cost of using FacTool in USD.
• Figure 3: An annotated FacTool-QA example sampled from FactBench.
• Figure 4: An annotated HaluEval example sampled from FactBench.
• Figure 5: The number of extracted atomic claims using FacTool across responses of LLaMA-2 7B, 13B, and GPT-4.