license: cc-by-4.0
task_categories:

• text-classification
  language:
• en
  tags:
• safety
• guardrails
• content-moderation
• jailbreak
• llm
• classification
  size_categories:
• 10K<n<100K
  configs:
• config_name: expguardtrain
  data_files:
  • split: train
    path: expguardtrain.parquet
• config_name: expguardtest
  data_files:
  • split: test
    path: expguardtest.parquet

extra_gated_prompt: >-
Access to this dataset is granted upon agreement to the terms below. This dataset is a research artifact for ExpGuard, containing examples of harmful content for the purpose of safety research.
extra_gated_fields:
Name: text
Organization or Affiliation: text
Country: text
Email: text
Intended use of the dataset: text
I understand that this dataset is a research artifact that may contain unfiltered, toxic, or harmful material: checkbox
I agree to use this dataset solely for research purposes to advance AI safety and not for malicious use: checkbox
I certify that the information I have provided is true and accurate: checkbox

Dataset Card for ExpGuardMix

Dataset Summary

ExpGuardMix is a comprehensive dataset designed for training and evaluating guardrail models, specifically tailored for specialized domains like finance, healthcare, and law. It consists of two splits: ExpGuardTrain and ExpGuardTest.

ExpGuardTrain (used to train ExpGuard):

• Data Size: 56,653 samples, consisting of 29,750 prompt-only examples and 26,903 prompt-response pairs.
• Composition:
  • Specialized Domains (19,907 samples): Finance, Healthcare, and Law. Constructed using a rigorous pipeline involving terminology mining from Wikipedia, filtering with Wikidata and human verification, and generation using GPT-4o.
  • General Domain (36,746 samples): "In-the-wild" user-LLM interactions and human-written data. Sources include LMSYS-Chat-1M, WildChat, Do-Anything-Now (DAN) jailbreaks, HH-RLHF, and Aegis 2.0.
• Labels: Binary labels for prompt harmfulness and response harmfulness. Labels were assigned by an ensemble of three state-of-the-art LLMs (Claude 3.7 Sonnet, Gemini 2.0 Flash, Qwen2.5-Max) using Chain-of-Thought reasoning, followed by strict majority voting.
• Note on Data Availability: WildChat, LMSYS-Chat-1M, and Suicide Detection subsets in Aegis 2.0 have been removed from the distributed version of this dataset. Users can access the subsets used in the paper through the ExpGuard GitHub repository.

ExpGuardTest (used to evaluate safety classifiers):

• Data Size: 2,275 high-quality samples.
• Domain Breakdown: 964 Financial, 771 Medical, and 540 Legal items.
• Labels: Expert-annotated. For the financial domain, labels were verified by domain experts (authors with banking experience). The test set shows high agreement between the ensemble and experts (Cohen's Kappa 0.89 for prompts, 0.98 for responses).

Please check the paper for further details on data construction: ExpGuard: LLM Content Moderation in Specialized Domains.

Risk Taxonomy

The dataset covers the following harm categories, inspired by the MLCommons Hazard Taxonomy and Weidinger et al. (2021) :

1. Violence & Incitement
2. Sexual Content & Exploitation
3. Hate Speech & Discrimination
4. Harassment & Bullying
5. Self-Harm & Suicide Promotion
6. Privacy Violation
7. Intellectual Property Infringement
8. Illegal Weapons
9. Controlled Substances
10. Criminal Planning
11. Fraud, Scams & Deception
12. Unqualified Professional Advice
13. Misinformation & Disinformation

Usage

from datasets import load_dataset

# Load the expguardtrain dataset
dataset = load_dataset("6rightjade/expguardmix", "expguardtrain")

# Load the expguardtest dataset
dataset = load_dataset("6rightjade/expguardmix", "expguardtest")

Dataset Details

The dataset contains the following columns:

• prompt: str, the user request.
• response: str, the model response (may be empty).
• prompt_label: str, safety label for the prompt (e.g., "safe", "unsafe").
• response_label: str, safety label for the response.
• prompt_category: str, specific harm category of the prompt (e.g., "Unharmful", "Fraud, Scams & Deception").
• response_category: str, specific harm category of the response.
• domain: str, the domain of the content (e.g., "finance", "medical", "legal").
• scenario: str, the specific scenario or topic.

Citation

@inproceedings{
  choi2026expguard,
  title={ExpGuard: {LLM} Content Moderation in Specialized Domains},
  author={Choi, Minseok and Kim, Dongjin and Yang, Seungbin and Kim, Subin and Kwak, Youngjun and Oh, Juyoung and Choo, Jaegul and Son, Jungmin},
  booktitle={The Fourteenth International Conference on Learning Representations},
  year={2026},
  url={https://openreview.net/forum?id=t5cYJlV6aJ}
}