HiTZ/ALIA_syntethic_MT_V2

ALIA Synthetic MT

ALIA Synthetic MT V2 is a parallel corpus derived from Berria news articles and legal/administrative domains BOPV and Parlamento, comprising content published in 2025 as well as archived material from 2023. The dataset provides synthetic translations into English and Spanish, generated using two distinct Large Language Models: Qwen3.5-27B and LatxaQ.

Model Details

This dataset utilizes the following models for translation generation:

1. Qwen3.5-27B: A state-of-the-art multilingual large language model utilized for the legal domain.
2. LatxaQ-VL: A domain-adapted iteration of Qwen3-32B utilized for news domain.

LatxaQ serves as a robust alternative to the standard Latxa model. It preserves the inherent Basque linguistic and cultural capabilities characteristic of the Latxa family while leveraging the advanced multilingual foundation of the Qwen architecture.

Dataset Structure

The dataset is formatted as a JSONL file. Each entry contains the following fields:

Berria

Paragraphs belonging to the same article are linked in order using sequential IDs (e.g., berria.202509-0-9363_0, berria.202509-0-9363_1), making it easy to reconstruct consecutive paragraphs whenever available. All translations in the dataset are aligned with a minimum similarity score of 0.75. We kept the exact score in the data so you can apply stricter filters if you want. Finally, if you see a null value in the doc_es field, it means the Spanish translation for that specific segment failed our quality checks and was removed.

• id: Unique identifier for the document.
• doc_eu: Original source text in Basque (Euskera).
• doc_en: Synthetic English translation.
• doc_es: Synthetic Spanish translation.
• sonar_en_sore: Sonar similarity score between the Basque and the English text.
• sonar_es_sore: Sonar similarity score between the Basque and the Spanish text.

BOVP and Parliament There are multiple metadata columns, the most important for MT task.

• eu_unique_id: Unique identifier for the basque document.
• es_unique_id: Unique identifier for the Spanish document.
• eu_text: Original source text in Basque (Euskera).
• es_text: Original source text in Spanish.
• translation: Synthetic English translation.

Statistics

Berria

• Total Documents in EU-ES: 337,650
• Total Documents in EU-ES-EN: 193,337
• Source: Berria (2023–2025)

BOPV

• Total Documents: 27,130
• Source: Boletín Oficial del País Vasco(2023–2025)

Parlamento

• Total Documents: 20,385
• Source: Parlamento(2023–2025)

Intended Use

This dataset is intended for:

• Training and evaluation of machine translation models
• Synthetic data experiments
• Multilingual NLP research Note: Since the translations are synthetic, they might contain model-specific artifacts, so results from training or evaluation should be interpreted with that in mind.

License

This data is released under Creative Commons CC-BY-SA license following the license of the original Berria, BOPV and Parliament data. Synthetic translations inherit the same usage restrictions as the source data.

Contact

For questions or issues, please contact the ALIA / HiTZ team or open an issue on the Hugging Face dataset repository.

Citation

If you use this dataset in your work, please cite the ALIA project and the corresponding translation models used.

Funding

This work is funded by the Ministerio para la Transformación Digital y de la Función Pública - Funded by EU – NextGenerationEU within the framework of the project Desarrollo de Modelos ALIA.