AutoHall: Automated Factuality Hallucination Dataset Generation for Large Language Models

TL;DR

AutoHall introduces an automated pipeline to generate model-specific factuality hallucination datasets by leveraging public fact-checking corpora, eliminating manual annotation. It couples reference generation, claim verification, and balanced hallucination collection, then pairs this with a zero-resource self-contradiction detector to identify hallucinations without external knowledge bases. Across multiple open- and closed-source LLMs, AutoHall reveals model- and topic-dependent hallucination patterns (roughly 15–30%), and its detector consistently outperforms zero-resource baselines in accuracy and F1. The work also provides extensive ablations and human validation, positioning AutoHall as a practical baseline for automated dataset creation and evaluation of hallucination detection in evolving LLMs.

Abstract

Large language models (LLMs) have gained broad applications across various domains but still struggle with hallucinations. Currently, hallucinations occur frequently in the generation of factual content and pose a great challenge to trustworthy LLMs. However, hallucination detection is hindered by the laborious and expensive manual annotation of hallucinatory content. Meanwhile, as different LLMs exhibit distinct types and rates of hallucination, the collection of hallucination datasets is inherently model-specific, which also increases the cost. To address this issue, this paper proposes a method called $\textbf{AutoHall}$ for $\underline{Auto}$matically constructing model-specific $\underline{Hall}$ucination datasets based on existing fact-checking datasets. The empirical results reveal variations in hallucination proportions and types among different models. Moreover, we introduce a zero-resource and black-box hallucination detection method based on self-contradiction to recognize the hallucination in our constructed dataset, achieving superior detection performance compared to baselines. Further analysis on our dataset provides insight into factors that may contribute to LLM hallucinations. Our codes and datasets are publicly available at https://github.com/zouyingcao/AutoHall.

Figures (6)

• Figure 1: A hallucination example. The red underline indicates the hallucinatory content generated by ChatGPT, since the novel never mentions the presence of a "red diamond" at the crime scene and the "The Snowman" case has also been solved before.
• Figure 2: Our proposed approach to collect LLM hallucination automatically. The grounded information is colored green. The incorrect information is colored red. Some analysis on prompt sensitivity is included in Section \ref{['sec:promptsensitivity']}.
• Figure 3: Our proposed approach to detect LLM hallucination. The claim from fact-checking dataset is colored blue. The response need to be detected whether exists hallucination is colored red. The sampled references to trigger self-contradictions are colored purple. The complete Step 2 prompts are shown in Tab. \ref{['tab:designed_prompts']}.
• Figure 4: Histogram for Num${_c}$ in hallucinatory and factual references (model: ChatGPT, TEMP: 0.1, dataset: WICE).
• Figure 5: Hallucination proportion across top 10 topics for GPT-4o, ChatGPT and Llama2-7B-Chat.