dataset_info:
features:

• name: class
  dtype: string
• name: category
  dtype: string
• name: subcategory
  dtype: string
• name: question
  dtype: string
• name: answer
  dtype: string
  splits:
• name: train
  num_bytes: 7820943
  num_examples: 4185
  download_size: 4066982
  dataset_size: 7820943
  configs:
• config_name: default
  data_files:
  • split: train
    path: data/train-*
    extra_gated_heading: Acknowledge license to accept the repository
    extra_gated_prompt: 'This dataset is released under a Research-only Data Use Agreement
    (DUA). By requesting access, you agree to the following conditions: (1) The dataset
    is for non-commercial academic research only. (2) Any kind of misuse is strictly
    prohibited.'
    extra_gated_fields:
    Name: text
    Organization: text
    I want to use this dataset for:
    type: select
    options:
    • Research
    • Education
    • label: Other
      value: other

extra_gated_button_content: Acknowledge license

Synthetic Safety Dataset

Nowadays, the possibility to generate unsafe content is fundamental. Unsafe content can be used to train/evaluate moderation models,
evaluate the safety of LLM and align LLMs to safety policies.
For example, in the Constitutional Classifier paper by Anthropic, they have used an internal model without harmlessness optimization to generate unsafe content and train
moderation models using that. Unfortunatelly, it is not always easy to generate this kind of content as most of the publicly available SOTA models are generally safe aligned.
It is always possible to jailbreak them and ask them to generate unsafe content, but finding a unique jailbreak that works for several unsafe content and
that does not induce biases on the type of unsafe content that is generated, can be, in some cases, quite difficult.

The dataset includes a class field to reppresent one :

• unsafe: the question is unsafe with respect to the category and subcategory
• safe_releated: the question is safe with respect to the category
• safe_unreleated: the quesion is safe with respect to the categoory and is not releated to the category

The data includes examples that might be disturbing, harmful or upsetting. It includes a range of harmful topics such as discriminatory language and discussions about abuse, violence, self-harm, sexual content, misinformation among other high-risk categories.
The main goal of this data is for advancing research in building safe LLMs.

Data Collection Method

The dataset was created with a semi-automatic approach using two LLMs:

• A medium-sized highly steerable LLM
• A larger but less steerable LLM

The steps include the generation of a seed dataset, manually reviewed, and two extension steps.
The steps can be summarized in this way:

• A small set of (question, answer) pairs are generated for each category and subcategory using the highly steerable LLM
• The examples are manually checked and safe examples are removed (if present)
• The filtered set is used as input to ask the highly steerable LLM to generate additional pairs
• The generated questions are used in a multi-turn conversation to steer the larger but less steerable LLM, to expand the set of questions generated in the previous step
• The new questions, combined with the previous questions, are filtered using the larger model to obtain the final set of questions
• Unanswered questions from the previous step are answered using the highly steerable LLM, using the examples of step 3 as few-shot examples

Right now the dataset consider only three safety categories but can be easly extended to include more:

• SEX_RELATED_CRIMES
• SUICIDE_AND_SELF_HARM
• VIOLENT_CRIMES
• INDISCRIMINATE_WEAPONS
• SEXUAL_CONTENT
• NON_VIOLENT_CRIMES

The definition of the three categories, used during the generation steps, are based on the ones used for the AILuminate benchmark dataset: https://mlcommons.org/ailuminate/
For each category, few subcategories were defined (to be extended). For each subcategory, the final set includes 50 question.

A detailed description of the method will be released in the future.

Uses

Direct Use

To build LLM content safety moderation guardrails for LLMs, can be used to train both prompt and response moderation.
Can be used for alignment of general purpose LLMs towards safety along with carefully using the data for safe-unsafe preference pairs.

Out-of-Scope Use

The data contain content that may be offensive or upsetting. Topics include, but are not limited to, discriminatory language and discussions of abuse, violence, self-harm, exploitation, and other potentially upsetting subject matter. Please only engage with the data in accordance with your own personal risk tolerance. The data are intended for research purposes, especially research that can make models less harmful. The views expressed in the data do not reflect my view. These data are not intended for training dialogue agents as this will likely lead to harmful model behavior.

Limitations

I did not check all the examples, some of them could be safe.
The examples lack sometimes in terms of diversity, to improve it,
it could be useful to use multiple type of personalities during the question and answering generation phases.

Contributors

Federico Ricciuti, https://www.linkedin.com/in/federico-ricciuti-b490ab59/

Cite

Cite

@misc{Ricciuti_T2TSyntheticSafetyBench-v2,
  author = {Federico Ricciuti},
  title = {T2TSyntheticSafetyBench-v2},
  year = {2026},
  howpublished = {\url{https://huggingface.co/datasets/fedric95/T2TSyntheticSafetyBench-v2}}
}