license: mit
language:
- en
tags: - safety
- red-teaming
- llm-evaluation
- refusal
- harmbench
pretty_name: HarmBench Scenarios
source_datasets: - walledai/HarmBench
configs: - config_name: en
data_files:- split: test
path: harmbench.en.jsonl
- split: test
HarmBench Scenarios
Safety-evaluation scenarios derived from the HarmBench behavior dataset, serialized as giskard Scenario objects (one JSON object per line).
Each row poses a harmful request the agent should refuse or safely decline, paired with an LLMJudge check (giskard.scan::judges/harmbench_safety.j2) that grades the agent response.
Format
Every line is a serialized giskard.checks.Scenario:
name—"HarmBench #<id>"steps[].interacts[].inputs— the harmful prompt sent to the agent under test (dataset_input). Contextual rows prepend the HarmBench context string using{context}\n\n---\n\n{behavior}.steps[].checks[].prompt_path—giskard.scan::judges/harmbench_safety.j2annotations—behavior,functional_category,semantic_category, optionalcontextandtagstags—dataset:harmbench,threat-type:harmful-content-generation, plus functional/semantic category tags
The en config has 400 scenarios in a single test split (evaluation-only, no training data):
| Functional category | Count |
|---|---|
| standard | 200 |
| contextual | 100 |
| copyright | 100 |
License & Attribution
This derivative is released under MIT and may be used commercially.
The source behavior prompts come from walledai/HarmBench (MIT). Please credit the original authors using the citation below.
Citation
@article{mazeika2024harmbench,
title={HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal},
author={Mantas Mazeika and Long Phan and Xuwang Yin and Andy Zou and Zifan Wang and Norman Mu and Elham Sakhaee and Nathaniel Li and Steven Basart and Bo Li and David Forsyth and Dan Hendrycks},
year={2024},
eprint={2402.04249},
archivePrefix={arXiv},
primaryClass={cs.LG}
}