license: cc-by-nc-sa-4.0
language:

• en
  tags:
• safety
• red-teaming
• llm-evaluation
• refusal
  pretty_name: Do-Not-Answer Scenarios
  source_datasets:
• LibrAI/do-not-answer
  configs:
• config_name: en
  data_files:
  • split: test
    path: donotanswer.en.jsonl

Do-Not-Answer Scenarios

Refusal-evaluation scenarios derived from the Do-Not-Answer dataset, serialized as
giskard Scenario objects (one JSON object per line).

Each row poses a question that a well-aligned assistant should decline, paired with a
Conformity check asserting the agent refused or safely declined.

Format

Every line is a serialized giskard.checks.Scenario:

• name — "Do not answer #<id>"
• steps[].interacts[].inputs — the question to send to the agent under test
• steps[].checks[].rule — refusal rule referencing the specific harm
• annotations — types_of_harm, risk_area, specific_harms (from the source dataset)
• tags — dataset:do-not-answer, threat-type:harmful-content-generation

The en config has 939 scenarios spanning 5 risk areas and 12 harm types,
in a single test split (evaluation-only, no training data).

License & Attribution

This derivative is released under CC-BY-NC-SA-4.0, inherited from the source dataset
(LibrAI/do-not-answer). The original
Do-Not-Answer source code is Apache-2.0; the data is CC-BY-NC-SA-4.0. Please credit the
original authors using the citation below.

Citation

@misc{wang2023donotanswer,
    author = {Wang, Yuxia and Li, Haonan and Han, Xudong and Nakov, Preslav and Baldwin, Timothy},
    title = {Do-Not-Answer: A Dataset for Evaluating Safeguards in LLMs},
    year = {2023},
    howpublished = {arXiv preprint arXiv:2308.13387},
}