dataset_info:
features:

• name: id
  dtype: int64
• name: seed_id
  dtype: string
• name: seed_text
  dtype: string
• name: language
  dtype: string
• name: group
  dtype: string
• name: scenario_category
  dtype: string
• name: scenario
  dtype: string
• name: jb_strategy
  dtype: string
• name: jb_prompt
  dtype: string
  splits:
• name: train
  num_bytes: 82072950
  num_examples: 43961
  download_size: 46938616
  dataset_size: 82072950
  configs:
• config_name: default
  data_files:
  • split: train
    path: data/train-*
    task_categories:
• text-generation
• text-classification
  language:
• en
• pt
  tags:
• llm-safety
• jailbreak
• bias
  pretty_name: MiJaBench
  size_categories:
• 10K<n<100K

Safety Is Not Universal: The Selective Safety Trap in LLM Alignment

Accepted to Findings of ACL 2026

Overview

MiJaBench: Minority Jailbreaking Benchmark

MiJaBench is a bilingual English-Portuguese adversarial benchmark for auditing demographic-specific vulnerabilities in large language model safety alignment. It was introduced in the paper:

Content Warning

This dataset contains adversarial prompts derived from toxic, hateful, and discriminatory content. It is intended strictly for safety evaluation, red teaming, fairness auditing, and alignment research. The dataset should not be used to generate harmful content, target protected groups, bypass safety mechanisms, or train malicious systems.

Dataset Summary

MiJaBench is designed to test whether LLM safety alignment protects different demographic groups consistently. Instead of reporting only aggregate scores for broad categories such as Identity Hate, MiJaBench enables fine-grained analysis of how model refusals vary across specific minority groups.

The benchmark contains 43,961 controlled jailbreak prompts in English and Portuguese. Each prompt is generated from three controlled components:

1. A hate-speech seed defining the harmful intent.
2. A contextual scenario that places the harmful intent in a varied narrative setting.
3. A jailbreaking strategy that defines the adversarial attack style.

This structure makes it possible to evaluate whether model safety behavior transfers across demographics, languages, scenarios, and attack strategies.

Dataset Structure

Data Instances

A typical instance contains the original toxic seed, the target group, the contextual scenario, the jailbreak strategy, and the final adversarial prompt.

Example structure:

{
  "id": "0",
  "seed_id": "toxigen_17193",
  "seed_text": "[toxic seed text]",
  "language": "english",
  "group": "asian",
  "scenario_category": "humor_and_absurdity",
  "scenario": "[contextual scenario text]",
  "jb_strategy": "representation_shifting",
  "jb_prompt": "[final adversarial jailbreak prompt]"
}

The actual dataset contains harmful and offensive text by design. The example above is partially sanitized.

Data Fields

• id (int): Identifier of the sample seed.
• seed_id (string): Identifier of the original seed. It combines the source seed dataset name with the original row ID. If the source dataset does not provide a row ID, the row index is used.
• seed_text (string): The original hate-speech seed used to define the harmful intent of the prompt.
• language (string): Language of the sample. The dataset contains English and Portuguese instances.
• group (string): The target demographic group associated with the seed and jailbreak prompt.
• scenario_category (string): The broad category of the contextual scenario.
• scenario (string): The specific contextual scenario used to condition the adversarial rewrite.
• jb_strategy (string): The jailbreak strategy used to structure the prompt.
• jb_prompt (string): The final adversarial jailbreak prompt generated from the seed, scenario, and strategy.

Dataset Statistics

Demographic coverage

Dataset Creation

MiJaBench was created through a controllable synthetic generation pipeline. First, hate-speech seeds were sampled from English and Portuguese hate-speech resources. Then, contextual scenarios were generated to increase linguistic and narrative diversity. Finally, each seed and scenario pair was rewritten using a controlled jailbreak strategy.

The final prompt, stored in jb_prompt, preserves the harmful intent of the original seed_text while varying the scenario and adversarial structure. This enables controlled analysis of safety failures across demographic groups, languages, scenarios, and jailbreak strategies.

Intended Uses

MiJaBench is intended for:

• Evaluating LLM safety alignment.
• Auditing demographic disparities in refusal behavior.
• Red-teaming models against identity-targeted harmful requests.
• Studying cross-lingual safety transfer between English and Portuguese.
• Developing safer and more equitable alignment methods.

MiJaBench should not be used to produce, amplify, or deploy hateful or discriminatory content.

Related Dataset: MiJaBench-Align

The companion dataset, MiJaBench-Align, contains model responses generated by evaluating MiJaBench across multiple LLMs. Each sample includes the original benchmark metadata, the evaluated model, the model response, and a safety verdict.

MiJaBench-Align contains 615,454 prompt-response pairs from 14 evaluated models.

Dataset: AKCIT/mijabench_align

MiJaBench contains adversarial prompts derived from toxic, hateful, and discriminatory content. These data are released exclusively for:

• Safety evaluation
• Red teaming
• Fairness auditing
• Alignment research

The benchmark should not be used to generate harmful content, target protected groups, bypass safety mechanisms, or develop malicious systems.

Citation

If you find this repository useful for your research or build upon our work, please cite:

@inproceedings{brito2026safety,
  title={Safety Is Not Universal: The Selective Safety Trap in LLM Alignment},
  author={Brito, Iago Alves and Rios, Walcy and Dollis, Julia Soares and Silva, Diogo Fernandes Costa and Galv{\~a}o Filho, Arlindo Rodrigues},
  booktitle={Findings of the Association for Computational Linguistics: ACL 2026},
  pages={10044--10065},
  year={2026}
}

Contact

For questions about the dataset, please open an issue in the GitHub repository or contact:

Iago Alves Brito
Email: [email protected]
GitHub: github.com/iagoalvesb/mijabench