AnnikaSimonsen/TrustLLM-reformulation-prompts

The TrustLLM Prompt Reformulation Dataset: A Germanic Languages Instruction Dataset

Dataset Description

This dataset contains culturally-aware multilingual instruction data for 8 Germanic languages, created using a novel prompt reformulation methodology. Rather than relying on direct translation, native speakers reformulated existing English prompts from OpenAssistant to ensure cultural authenticity and linguistic appropriateness.

Supported Languages

• Icelandic: 998 samples (12 contributors)
• Faroese: 997 samples (4 contributors)
• German: 250 samples (9 contributors)
• Danish: 999 samples (6 contributors)
• Dutch: 697 samples (11 contributors)
• Norwegian (Bokmål): 999 samples (1 contributors)
• Norwegian (Nynorsk): 998 samples (1 contributors)
• Swedish: 999 samples (1 contributors)

Dataset Statistics

• Total samples: 6,937
• Languages: 8
• Total contributors: 43
• Multilingual contributors: 2

Sample Types

• original: 3,987 samples
• reformulated: 1,495 samples
• reformulation: 1,455 samples

Methodology

This dataset was created using a prompt reformulation approach that:

1. Preserves task diversity while ensuring cultural and linguistic appropriateness
2. Combines human-reformulated prompts with synthetic responses for cost-effectiveness
3. Emphasizes cultural adaptation over direct translation
4. Leverages native speaker knowledge to create natural, contextually appropriate content

Reformulation Guidelines

Contributors were asked to:

• Replace English-specific cultural references with locally relevant ones
• Use idiomatic expressions and natural discourse patterns
• Adapt examples and scenarios to reflect local contexts
• Maintain core instructional intent while allowing semantic flexibility

Dataset Structure

Data Fields

• id: Unique identifier for each sample
• original_text: The instruction/prompt text in the target language
• language: Target language of the sample
• created_by: Contributor identifier (anonymized)
• created_at: Timestamp of creation
• type: Sample type (original vs reformulated)

Data Splits

This dataset does not have predefined train/validation/test splits. Users should create appropriate splits based on their specific use case.

Usage

from datasets import load_dataset

dataset = load_dataset("AnnikaSimonsen/TrustLLM-reformulation-prompts")

# Load specific language
danish_data = dataset.filter(lambda x: x['language'] == 'Danish')

# Load only original prompts
original_prompts = dataset.filter(lambda x: x['type'] == 'original')

Language-Specific Details

Icelandic

• Total samples: 998
• Contributors: 12
• Sample breakdown:
  • reformulated: 418
  • reformulation: 387
  • original: 193

Faroese

• Total samples: 997
• Contributors: 4
• Sample breakdown:
  • reformulation: 359
  • reformulated: 346
  • original: 292

German

• Total samples: 250
• Contributors: 9
• Sample breakdown:
  • reformulated: 125
  • reformulation: 121
  • original: 4

Danish

• Total samples: 999
• Contributors: 6
• Sample breakdown:
  • reformulated: 388
  • reformulation: 339
  • original: 272

Dutch

• Total samples: 697
• Contributors: 11
• Sample breakdown:
  • reformulation: 249
  • original: 230
  • reformulated: 218

Norwegian (Bokmål)

• Total samples: 999
• Contributors: 1
• Sample breakdown:
  • original: 999

Norwegian (Nynorsk)

• Total samples: 998
• Contributors: 1
• Sample breakdown:
  • original: 998

Swedish

• Total samples: 999
• Contributors: 1
• Sample breakdown:
  • original: 999

Ethical Considerations

This dataset was created with explicit attention to cultural sensitivity and linguistic authenticity. The reformulation methodology ensures that:

• Cultural references are appropriate for target language communities
• Linguistic patterns reflect natural usage rather than translation artifacts
• Content respects local contexts and cultural norms

Citation

If you use this dataset, please cite our paper:

@inproceedings{simonsen2022reformulate,
  title={Reformulate and Create, Don't Translate: Creating Natural Prompts for Underserved Languages},
  author={Simonsen, Annika and Stenlund, Hans Erik Mathias and Bungum, Lars and Daníel Skipstað Volhardt, Marc and Einarsson, Hafsteinn},
  booktitle={Proceedings of LREC 20225},
  year={2025},
  address={Palma de Mallorca, Spain}
}

Dataset Creators

• Annika Simonsen (University of Iceland)
• Hans Erik Mathias Stenlund (University of Iceland)
• Lars Bungum (Norwegian University of Science and Technology)
• Hafsteinn Einarsson (University of Iceland)

Acknowledgements

We thank all community contributors who participated in prompt reformulation and our collaborators for valuable feedback. Special recognition goes to the dedicated TrustLLM members who contributed their time to this effort and the Open Assistant project for sharing their crowd sourcing data openly online.

License

This dataset is released under the Creative Commons Attribution 4.0 International License (CC BY 4.0).