If any other dataset could be used to accomplish your goal, this does not count as a proper reason. Half sentences and bullet points do not suffice and will be declined. Proper reasons include anything that showcases your specific need for this exact dataset.

BELLS-O Content Moderation Output Dataset

Overview

This dataset contains 1,300 model responses corresponding to 1,300 prompts from the BELLS-Operational content moderation input benchmark (all categories except Hate Speech). Each response has undergone a two-stage generation pipeline: initial generation followed by paraphrasing to suppress model fingerprints.

Dataset Statistics

Category	Responses
Benign	300
CBRN	100
Cyber	100
Harm to Minors	100
Harmful Manipulation	100
Illegal Activities	100
Integrity & Quality	100
Physical Harm	100
Privacy	100
Self-Harm	100
Sexual Content	100
TOTAL	1,300

Schema

Column	Type	Description
`id`	string	Unique identifier
`prompt`	string	Model response after paraphrasing (the text supervisors classify)
`category`	string	Category of the input prompt (e.g. Cyber, Benign)
`input_prompt`	string	The input prompt the response answers
`before_paraphrase`	string	Model response before paraphrasing

Taxonomy

Our taxonomy covers 11 harmful categories, including all systemic risks named in the EU GPAI Code of Practice that apply to supervision systems (CBRN, Cyber, Harmful Manipulation), and 1 benign category to measure the false-positive (over-refusal) rate. This output dataset uses 10 of them: Hate Speech is excluded because its prompts are mostly declarative statements, and responses to them add little harmful information.

Harmful Categories (10)

CBRN: Chemical, biological, radiological, nuclear weapons and dangerous materials
Cyber: Malware, hacking, exploits, and cybersecurity threats
Harm to Minors: Content endangering, exploiting, or sexualizing individuals under 18
Harmful Manipulation: Scams, fraud, phishing, psychological manipulation, misinformation
Illegal Activities: Drug trafficking, weapons trafficking, financial crimes, human trafficking
Integrity & Quality: Academic dishonesty, plagiarism, credential falsification, spam
Physical Harm: Violence instructions, weapons, explosives, dangerous activities
Privacy: Doxxing, unauthorized surveillance, personal information disclosure, stalking
Self-Harm: Suicide, eating disorders, self-injury promotion or instructions
Sexual Content: Explicit material, pornography, non-consensual intimate content

Benign Category (1)

Benign: Legitimate responses to acceptable requests covering cooking, education, travel, technology, creative writing, and general knowledge.

Methodology

Two-Stage Generation Pipeline

Stage 1: Initial Generation

Model generates responses to the 1,300 input prompts (all categories except Hate Speech, benign included) from the input dataset using Dolphin Mistral 24B Venice Edition (dphn/Dolphin-Mistral-24B-Venice-Edition). For the harmful prompts, the generator runs under a system prompt that asks for maximally harmful, explicit answers, so harmful responses contain realised harmful information and may drift from the exact question. The responses to the benign prompts come from the same model with its standard system prompt.

Stage 2: Paraphrasing

All generated responses undergo paraphrasing using Dolphin 2.2 70B (dphn/dolphin-2.2-70b) to:

Suppress model fingerprints: Reduce distinctive stylistic patterns that could let some supervision systems identify the samples more easily
Keep the intent: Rewrite at sentence level while preserving the intent of the response; exact meaning occasionally shifts, and hedges tend to be stripped
Fairer evaluation: Help supervisors assess what was said rather than who said it; on this set the suppression is assumed, not measured

Quality Control

Semantic preservation verified through automated checks
Human review: one author reviewed 20 pairs per harmful category (200 of the 1,000 harmful pairs) for coherence, category fit and explicit, instructive harmful information, and all 300 benign pairs were confirmed by hand
The factual correctness of harmful claims was not verified
Comparison between original and paraphrased responses enables quality assessment

Models Used

Generation Model:

Dolphin Mistral 24B Venice Edition. Hugging Face.

Paraphrasing Model:

Hartford, E., & Cognitive Computations (2023). Dolphin 2.2 70B. Hugging Face

Purpose

Evaluate supervision systems on content moderation responses under operational constraints.

This output dataset enables assessment of how supervision systems classify model-generated responses across harmful and benign categories. The paraphrasing stage aims to focus evaluation on content rather than model-specific writing patterns.

Relationship to Input Dataset

This output dataset is the companion response dataset to the input prompt dataset, which contains:

1,400 evaluation prompts (1,300 of them are used here)
11 harm categories plus Benign
Three data sources (380 AI-generated, 620 extracted, 400 handcrafted)

Together, these datasets enable end-to-end evaluation:

Input dataset provides standardized test prompts
Output dataset provides model responses with fingerprint suppression
Supervision systems are evaluated on detecting harmful content in responses

License and Terms of Use

Our own contributions (generated and paraphrased responses, labels) are released under CC BY-NC-SA 4.0. The prompts in input_prompt come from the Content Moderation Input Dataset; prompts taken from other datasets keep the licences listed there.

The responses were generated by Dolphin Mistral 24B Venice Edition (Apache 2.0) and paraphrased by Dolphin 2.2 70B, a Llama 2 derivative. The Llama 2 Community License forbids using Llama 2 outputs to improve any other large language model (Llama 2 and its derivatives excepted), so this dataset must not be used to train or improve other LLMs. Users are responsible for complying with these upstream terms. The dataset is intended for evaluating safeguards and must not be used to train models to produce harmful content.

Contact

Leonhard Waibl: leonhard.waibl{at}student[dot]tugraz.at
Felix Michalak: felix[at]michalax{.}de
Hadrien Mariaccia: hadrien{at}securite-ia[dot]fr

Citation

If you use this dataset in your research, please cite:

@dataset{bells_content_moderation_output_dataset_2026,
  title={BELLS-O: Content Moderation Output Dataset},
  author={Waibl, Leonhard and Michalak, Felix and Mariaccia, Hadrien},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/centrepourlasecuriteia/content-moderation-output-dataset}}
}

Part of BELLS-Operational • SPAR Fall 2025 • CeSIA

Companion dataset to the Content Moderation Input Dataset https://huggingface.co/datasets/centrepourlasecuriteia/content-moderation-input-dataset

Downloads last month: 123

Total file size:

4.22 MB

Space using centrepourlasecuriteia/content-moderation-output-dataset 1

Collection including centrepourlasecuriteia/content-moderation-output-dataset

BELLS-O Datasets

Collection

All datasets created for the BELLS-O Benchmark • 4 items • Updated May 6