Datasets:

QCRI
/

FanarGuard

language:
  - ar
  - en
license: other
task_categories:
  - text-classification
  - text-generation
tags:
  - cultural-safety
  - arabic
  - moderation
  - safety
  - middle-east
  - cultural-alignment
pretty_name: Cultural Safety Dataset
size_categories:
  - 1K<n<10K
dataset_info:
  features:
    - name: Response
      dtype: string
    - name: Prompt
      dtype: string
    - name: Model
      dtype: string
    - name: Judge_Qwen2.5-72B-Instruct
      dtype: string
    - name: Judge_Qwen3-32B
      dtype: string
    - name: Judge_gemma-2-27b-it
      dtype: string
    - name: Judge_c4ai-command-r-plus
      dtype: string
    - name: Qwen2.5-72B-Instruct_score
      dtype: int64
    - name: Qwen3-32B_score
      dtype: int64
    - name: gemma-2-27b-it_score
      dtype: float64
    - name: c4ai-command-r-plus_score
      dtype: float64
    - name: FanarGuard-R
      list: float64
    - name: FanarGuard-G-2B
      list: float64
    - name: FanarGuard-G-4B
      list: float64
    - name: Ann_1_score
      dtype: string
    - name: Ann_2_score
      dtype: string
    - name: Ann_3_score
      dtype: string
    - name: Judge_Average
      dtype: float64
    - name: Ann_Average
      dtype: float64
    - name: Taxonomy
      dtype: string
    - name: Data Source
      dtype: string
    - name: PAM
      dtype: bool
  splits:
    - name: train
      num_bytes: 11833983
      num_examples: 1451
  download_size: 5304127
  dataset_size: 11833983
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
extra_gated_fields:
  Full name: text
  Institutional affiliation: text
  Country you are located in: country
  Contact email: text
  I want to use this dataset for: text
  I agree to use this dataset exclusively for research purposes: checkbox
  I agree that I will not use this dataset for malicious purposes, including training models to generate harmful content or automating policy evasion: checkbox
  I agree that the dataset creators and their affiliated institutions are not liable for any claims, damages, algorithmic failures, or reputational harm resulting from my use or interpretation of this data: checkbox
  I certify that the information I have provided is true and accurate: checkbox

Cultural Safety Dataset

Dataset Description

The Cultural Safety Dataset is a benchmark for evaluating culturally sensitive and culturally misaligned model outputs in Arabic and Middle Eastern contexts. It focuses on cases where model responses conflict with culturally dependent societal norms and values.

The dataset was developed as part of FanarGuard: a culturally-aware moderation filter for Arabic language models.

Dataset Construction

The dataset combines:

822 prompts identified from production logs of an Arabic-language chat interface.
84 regionally sensitive questions from the Arabic Safety Benchmark.
198 manually generated prompts.

Three bilingual (English–Arabic) annotators classified the prompts for cultural relevance. The final set contains:

Category	Number
Culturally dependent	1,008
Partially cultural	36
General safety	60

The 1,008 culturally dependent prompts cover eight categories:

Family & Social Norms
Gender Roles & Equality
Health & Bodily Autonomy
Legal & Governance Norms
Identity & Minority Representation
Sexuality & Gender Identity
Political & Geopolitical Sensitivity
Religious Insult & Blasphemy

Model Responses

Responses were generated using five models:

GPT-4o
Qwen-3-32B
Gemma-3-27B-It
Fanar-1-9B-Instruct
ALLaM-7B-Instruct-Preview

The benchmark contains 1,451 question–answer pairs, which were evaluated by three bilingual annotators. 363 responses received a score below 3, indicating cultural misalignment.

Intended Use

This dataset is intended for:

Evaluating culturally aware moderation filters
Benchmarking Arabic language models
Studying cultural alignment and safety
Developing culturally informed safety classifiers

Limitations

The dataset focuses on Arabic and Middle Eastern contexts and does not represent all Arabic-speaking communities or cultural perspectives. Cultural norms vary across countries, communities, and individuals, and human annotations may involve subjective judgments.

Citation

If you use this dataset, please cite:

@inproceedings{fatehkia2026fanarguard,
  title={FanarGuard: a culturally-aware moderation filter for Arabic language models},
  author={Fatehkia, Masoomali and Altinisik, Enes and Sencar, Husrev Taha},
  booktitle={Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)},
  pages={7848--7869},
  year={2026}
}