language:

• pl
• en
• cs
• sk
• sl
• lt
• lv
• de
• it
• fr
• es
• pt
  multilinguality: multilingual
  pretty_name: Multilingual Red-Teaming Preference + Evaluation
  size_categories:
• 100K<n<1M
  task_categories:
• text-generation
• text-classification
  tags:
• safety
• red-teaming
• multilingual
• preference-dataset
• evaluation
  configs:
• config_name: evaluation
  data_files:
  • split: eval
    path: evaluation/eval-*
• config_name: lang_cs
  data_files:
  • split: train
    path: lang_cs/train-*
  • split: test
    path: lang_cs/test-*
• config_name: lang_de
  data_files:
  • split: train
    path: lang_de/train-*
  • split: test
    path: lang_de/test-*
• config_name: lang_en
  data_files:
  • split: train
    path: lang_en/train-*
  • split: test
    path: lang_en/test-*
• config_name: lang_es
  data_files:
  • split: train
    path: lang_es/train-*
  • split: test
    path: lang_es/test-*
• config_name: lang_fr
  data_files:
  • split: train
    path: lang_fr/train-*
  • split: test
    path: lang_fr/test-*
• config_name: lang_it
  data_files:
  • split: train
    path: lang_it/train-*
  • split: test
    path: lang_it/test-*
• config_name: lang_lt
  data_files:
  • split: train
    path: lang_lt/train-*
  • split: test
    path: lang_lt/test-*
• config_name: lang_lv
  data_files:
  • split: train
    path: lang_lv/train-*
  • split: test
    path: lang_lv/test-*
• config_name: lang_pl
  data_files:
  • split: train
    path: lang_pl/train-*
  • split: test
    path: lang_pl/test-*
• config_name: lang_pt
  data_files:
  • split: train
    path: lang_pt/train-*
  • split: test
    path: lang_pt/test-*
• config_name: lang_sk
  data_files:
  • split: train
    path: lang_sk/train-*
  • split: test
    path: lang_sk/test-*
• config_name: lang_sl
  data_files:
  • split: train
    path: lang_sl/train-*
  • split: test
    path: lang_sl/test-*
• config_name: multilingual
  data_files:
  • split: train
    path: multilingual/train-*
  • split: test
    path: multilingual/test-*
    dataset_info:
• config_name: evaluation
  features:
  • name: row_id
    dtype: int64
  • name: lang
    dtype: string
  • name: prompt
    dtype: string
    splits:
  • name: eval
    num_bytes: 5805282
    num_examples: 16800
    download_size: 3416852
    dataset_size: 5805282
• config_name: lang_cs
  features:
  • name: chosen
    list:
    • name: content
      dtype: string
    • name: role
      dtype: string
  • name: rejected
    list:
    • name: content
      dtype: string
    • name: role
      dtype: string
  • name: lang
    dtype: string
  • name: id
    dtype: string
  • name: row_id
    dtype: int64
  • name: category
    dtype: string
    splits:
  • name: train
    num_bytes: 12671899
    num_examples: 2748
  • name: test
    num_bytes: 654759
    num_examples: 141
    download_size: 5880244
    dataset_size: 13326658
• config_name: lang_de
  features:
  • name: chosen
    list:
    • name: content
      dtype: string
    • name: role
      dtype: string
  • name: rejected
    list:
    • name: content
      dtype: string
    • name: role
      dtype: string
  • name: lang
    dtype: string
  • name: id
    dtype: string
  • name: row_id
    dtype: int64
  • name: category
    dtype: string
    splits:
  • name: train
    num_bytes: 12135732
    num_examples: 2748
  • name: test
    num_bytes: 630927
    num_examples: 141
    download_size: 5373693
    dataset_size: 12766659
• config_name: lang_en
  features:
  • name: chosen
    list:
    • name: content
      dtype: string
    • name: role
      dtype: string
  • name: rejected
    list:
    • name: content
      dtype: string
    • name: role
      dtype: string
  • name: lang
    dtype: string
  • name: id
    dtype: string
  • name: row_id
    dtype: int64
  • name: category
    dtype: string
    splits:
  • name: train
    num_bytes: 13198313
    num_examples: 2748
  • name: test
    num_bytes: 663494
    num_examples: 141
    download_size: 6325215
    dataset_size: 13861807
• config_name: lang_es
  features:
  • name: chosen
    list:
    • name: content
      dtype: string
    • name: role
      dtype: string
  • name: rejected
    list:
    • name: content
      dtype: string
    • name: role
      dtype: string
  • name: lang
    dtype: string
  • name: id
    dtype: string
  • name: row_id
    dtype: int64
  • name: category
    dtype: string
    splits:
  • name: train
    num_bytes: 14577097
    num_examples: 2748
  • name: test
    num_bytes: 731766
    num_examples: 141
    download_size: 6323679
    dataset_size: 15308863
• config_name: lang_fr
  features:
  • name: chosen
    list:
    • name: content
      dtype: string
    • name: role
      dtype: string
  • name: rejected
    list:
    • name: content
      dtype: string
    • name: role
      dtype: string
  • name: lang
    dtype: string
  • name: id
    dtype: string
  • name: row_id
    dtype: int64
  • name: category
    dtype: string
    splits:
  • name: train
    num_bytes: 14429903
    num_examples: 2748
  • name: test
    num_bytes: 742599
    num_examples: 141
    download_size: 6196775
    dataset_size: 15172502
• config_name: lang_it
  features:
  • name: chosen
    list:
    • name: content
      dtype: string
    • name: role
      dtype: string
  • name: rejected
    list:
    • name: content
      dtype: string
    • name: role
      dtype: string
  • name: lang
    dtype: string
  • name: id
    dtype: string
  • name: row_id
    dtype: int64
  • name: category
    dtype: string
    splits:
  • name: train
    num_bytes: 14361447
    num_examples: 2748
  • name: test
    num_bytes: 723617
    num_examples: 141
    download_size: 6390570
    dataset_size: 15085064
• config_name: lang_lt
  features:
  • name: chosen
    list:
    • name: content
      dtype: string
    • name: role
      dtype: string
  • name: rejected
    list:
    • name: content
      dtype: string
    • name: role
      dtype: string
  • name: lang
    dtype: string
  • name: id
    dtype: string
  • name: row_id
    dtype: int64
  • name: category
    dtype: string
    splits:
  • name: train
    num_bytes: 10120891
    num_examples: 2748
  • name: test
    num_bytes: 532227
    num_examples: 141
    download_size: 4593588
    dataset_size: 10653118
• config_name: lang_lv
  features:
  • name: chosen
    list:
    • name: content
      dtype: string
    • name: role
      dtype: string
  • name: rejected
    list:
    • name: content
      dtype: string
    • name: role
      dtype: string
  • name: lang
    dtype: string
  • name: id
    dtype: string
  • name: row_id
    dtype: int64
  • name: category
    dtype: string
    splits:
  • name: train
    num_bytes: 12361380
    num_examples: 2748
  • name: test
    num_bytes: 636505
    num_examples: 141
    download_size: 5017826
    dataset_size: 12997885
• config_name: lang_pl
  features:
  • name: chosen
    list:
    • name: content
      dtype: string
    • name: role
      dtype: string
  • name: rejected
    list:
    • name: content
      dtype: string
    • name: role
      dtype: string
  • name: lang
    dtype: string
  • name: id
    dtype: string
  • name: row_id
    dtype: int64
  • name: category
    dtype: string
    splits:
  • name: train
    num_bytes: 11914709
    num_examples: 2748
  • name: test
    num_bytes: 610815
    num_examples: 141
    download_size: 5537258
    dataset_size: 12525524
• config_name: lang_pt
  features:
  • name: chosen
    list:
    • name: content
      dtype: string
    • name: role
      dtype: string
  • name: rejected
    list:
    • name: content
      dtype: string
    • name: role
      dtype: string
  • name: lang
    dtype: string
  • name: id
    dtype: string
  • name: row_id
    dtype: int64
  • name: category
    dtype: string
    splits:
  • name: train
    num_bytes: 13758159
    num_examples: 2748
  • name: test
    num_bytes: 710808
    num_examples: 141
    download_size: 6030431
    dataset_size: 14468967
• config_name: lang_sk
  features:
  • name: chosen
    list:
    • name: content
      dtype: string
    • name: role
      dtype: string
  • name: rejected
    list:
    • name: content
      dtype: string
    • name: role
      dtype: string
  • name: lang
    dtype: string
  • name: id
    dtype: string
  • name: row_id
    dtype: int64
  • name: category
    dtype: string
    splits:
  • name: train
    num_bytes: 11692081
    num_examples: 2748
  • name: test
    num_bytes: 598680
    num_examples: 141
    download_size: 5256586
    dataset_size: 12290761
• config_name: lang_sl
  features:
  • name: chosen
    list:
    • name: content
      dtype: string
    • name: role
      dtype: string
  • name: rejected
    list:
    • name: content
      dtype: string
    • name: role
      dtype: string
  • name: lang
    dtype: string
  • name: id
    dtype: string
  • name: row_id
    dtype: int64
  • name: category
    dtype: string
    splits:
  • name: train
    num_bytes: 10083361
    num_examples: 2748
  • name: test
    num_bytes: 515802
    num_examples: 141
    download_size: 4622500
    dataset_size: 10599163
• config_name: multilingual
  features:
  • name: chosen
    list:
    • name: content
      dtype: string
    • name: role
      dtype: string
  • name: rejected
    list:
    • name: content
      dtype: string
    • name: role
      dtype: string
  • name: lang
    dtype: string
  • name: id
    dtype: string
  • name: row_id
    dtype: int64
  • name: category
    dtype: string
    splits:
  • name: train
    num_bytes: 151304972
    num_examples: 32976
  • name: test
    num_bytes: 7751999
    num_examples: 1692
    download_size: 87562356
    dataset_size: 159056971
    license: llama3.1

Multilingual Red-Teaming Preference + Evaluation

This dataset contains multilingual red-teaming prompts and preference pairs for safety research, plus a separate multilingual evaluation prompt set.

Dataset Structure

The repository is organized into multiple dataset configs:

• multilingual
  • train: full preference training split
  • test: full preference test split
• lang_pl, lang_en, lang_cs, lang_sk, lang_sl, lang_lt, lang_lv, lang_de, lang_it, lang_fr, lang_es, lang_pt
  • train: language-specific preference training split
  • test: language-specific preference test split
• evaluation
  • eval: multilingual evaluation prompts

Data Fields

Config multilingual and lang_*

Each row is a preference sample with fields:

• chosen: list of chat turns (preferred response)
• rejected: list of chat turns (non-preferred response)
• lang: language code
• id: sample id
• row_id: original row id
• category: category label

Config evaluation

Each row is one prompt instance with fields:

• row_id: source CSV row index
• lang: language code
• prompt: evaluation prompt text

Intended Use

This dataset is intended for:

• safety alignment and preference learning research
• multilingual refusal and robustness benchmarking
• model evaluation under potentially harmful prompt distributions

Safety Notice

The dataset contains harmful, abusive, and policy-violating user prompts, and some rejected assistant responses may include unsafe content patterns. Handle with care and use only in controlled research environments.

Ethical Considerations

• Do not use this dataset to build systems that facilitate abuse.
• Apply strict filtering and governance for downstream usage.
• Ensure compliance with local legal and institutional requirements.

License

Specify the license used for this dataset before public release.

If unsure, keep the repository private until licensing and data governance are fully validated.

Quick Start

from datasets import load_dataset

repo_id = "NASK-PIB/RefusEU"

# Full multilingual preference dataset
train_ds = load_dataset(repo_id, "multilingual", split="train")
test_ds = load_dataset(repo_id, "multilingual", split="test")

# Language-specific subset
pl_train = load_dataset(repo_id, "lang_pl", split="train")
en_test = load_dataset(repo_id, "lang_en", split="test")

# Evaluation prompts
eval_ds = load_dataset(repo_id, "evaluation", split="eval")

📚 Citation

If you use the dataset, please cite the following paper:

@inproceedings{krasnodebska2026refuseu,
  author    = {Krasnodębska, Aleksandra and Kusa, Wojciech and Lipani, Aldo},
  title     = {{Multilingual Refusal Alignment for Safer Large Language Models}},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year      = {2026},
  address   = {San Diego, California, United States},
  publisher = {Association for Computational Linguistics}
}