By continuing, you confirm that the information below is accurate and that you will use this dataset only for controlled research or education, will not expose unsuspecting users to harmful content, and will not redistribute the files.

Hazardous Viewpoints Dataset (HVD)

HVD accompanies the EMNLP 2026 paper “Steering LLM Viewpoints through Fabricated Evidence Injection”. It is intended for controlled research on LLM safety, fabricated-evidence injection, viewpoint alignment, and defensive evaluation.

Official implementation: XiYang-ust/ghostwriter

Content warning: The dataset contains synthetic or LLM-generated stereotypes, misinformation, brand and political claims, unsafe medical claims, and psychologically harmful statements. Some entries may be offensive or distressing. None of the claims should be treated as factual guidance.

Dataset variants

HVD-G: 729 entries seeded with Grok and manually filtered. It contains 301 stereotype entries and 428 falsehood or manipulation entries. Every entry includes one explicit query and one implicit, decision-oriented query.
HVD-O: 729 parallel entries independently generated with an uncensored open-source model for generator-robustness analysis. HVD-O contains explicit queries; no separate implicit-query set was constructed for this variant.

Both variants span 11 categories: age, gender, occupation, physical characteristics, race, religion, sexual orientation, brand falsehoods, political views, medical falsehoods, and psychological manipulation.

Intended uses

evaluating whether LLMs adopt misleading viewpoints presented with fabricated evidence;
benchmarking detection and mitigation methods;
controlled reproducibility and robustness studies;
teaching AI-safety concepts in supervised settings.

Out-of-scope uses

user-facing deployment or live persuasion experiments;
targeting people or protected groups;
medical, psychological, political, or commercial decision support;
training systems to generate more convincing harmful content;
redistribution to people who have not independently requested access and accepted these terms.

Data fields

Both configurations contain:

id: unique record identifier within a variant;
variant: hvd_g or hvd_o;
category: one of the 11 categories above;
severity: source severity annotation where available;
basic_statement: the viewpoint or claim under evaluation;
explicit_query: a user query that directly raises the relevant topic;
explanation: why the statement is misleading, biased, or unsafe.

HVD-G additionally contains implicit_query: a plausible standalone decision, advice, or comparison query that can be influenced by the viewpoint without stating it directly. HVD-O does not contain this field because no corresponding implicit-query set was constructed.

The release intentionally excludes paraphrased explicit-query variants, unused scenario context and record-type labels, repackaged statements, target-model responses, automated scores, and experiment outputs.

Limitations and biases

HVD is generated and curated rather than sampled from natural conversations. It may reflect the biases, writing patterns, factual errors, and safety behavior of its generator models and annotators. Category coverage is not a prevalence estimate, and model performance on HVD should not be interpreted as real-world harm frequency. Entries involving brands, public figures, health, or psychology may be false by design.

No personal data is intentionally collected. Because synthetic text can accidentally reproduce real names or claims, users must avoid attempts to identify people and should report suspected personal information to the maintainers.

Access requests

Access is granted automatically after a signed-in Hugging Face user completes the request form and accepts the terms above; there is no project-by-project review. Hugging Face shares the requester's username and email address with the dataset maintainers, together with the additional fields in the form. This information is used to understand how the dataset is used, manage access, and communicate material safety or terms updates. Do not submit sensitive personal information in the free-text fields.

The request information is not included in the downloadable dataset files. Dataset maintainers may revoke access if use is inconsistent with these terms.

Research-use terms

By accessing the files, you agree to:

use the dataset only for the controlled research or educational purpose declared in the request form;
not expose unsuspecting users to dataset content or derived harmful generations;
not use it for harassment, discrimination, manipulation, fraud, unsafe advice, or influence operations;
not redistribute the dataset or share credentials; collaborators must request access independently and accept these terms;
secure local copies and delete them when the declared project ends;
cite the accompanying paper in publications using the dataset;
comply with applicable law, institutional review requirements, and model-provider terms.

Citation

@article{yang2026steering,
  title   = {Steering LLM Viewpoints through Fabricated Evidence Injection},
  author  = {Yang, Xi and Liu, Chang and Huang, Zhenglin and Li, Haoran and Zhang, Weiming and Weng, Jian and Song, Yangqiu},
  journal = {arXiv preprint arXiv:2606.06244},
  year    = {2026}
}

Downloads last month: -

Total file size:

838 kB

Paper for Promptist/HVD

Steering LLM Viewpoints through Fabricated Evidence Injection

Paper • 2606.06244 • Published Jun 4