yaojialzc/JailFlowEval

---
pretty_name: JailFlowEval
language:
- en
license: cc-by-4.0
size_categories:
- n<1K
task_categories:
- text-generation
tags:
- datasets
- benchmark
- safety
- red-teaming
- jailbreak
- agent-security
- workflow-safety
- enterprise-workflows
configs:
- config_name: default
data_files:
- split: train
path: data/train.parquet
---

# JailFlowEval

This repository contains the dataset subset for our paper **"JailFlowEval: A Workflow-Safety Benchmark for LLM-Driven Enterprise Service Workflows"**.

## Dataset Summary

- Rows: `744`
- Unique source prompts: `373`
- Split: `train`

Task distribution:

- `plan=336`
- `exec=202`
- `review=103`
- `improve=103`

## Fields

Each row contains:

- `example_id`: unique row identifier in this packaged dataset
- `source_id`: prompt ID from the source table
- `quality_tier`: always `high_quality`
- `task`: one of `plan`, `exec`, `review`, or `improve`
- `source`: source dataset name
- `cluster_label`: semantic cluster label
- `forbidden_prompt`: original harmful or disallowed prompt
- `jailbroken_prompt`: rewritten situational jailbreak prompt
- `response`: target model response
- `score`: PAIR-based harmfulness score

## Safety Notice

This dataset contains harmful prompts and harmful or policy-violating model outputs. It is intended for safety evaluation, red teaming research, and defense analysis. It should not be used to facilitate real-world harm.

## Citation

If you use this dataset, please cite this dataset release. The formal manuscript citation can be added after publication.

```bibtex
@misc{liu2026jailfloweval_dataset,
title = {JailFlowEval: A Workflow-Safety Benchmark for LLM-Driven Enterprise Service Workflows},
author = {Zhichao Liu and Haining Yu and Wenbo Pan and Wenting Zhang and Bingxuan Wang and Xiang Li},
year = {2026},
howpublished = {\url{https://huggingface.co/datasets/yaojialzc/JailFlowEval}},
note = {Dataset release}
}
```