metadata

dataset_info:
  features:
    - name: conversations
      list:
        - name: from
          dtype: string
        - name: value
          dtype: string
    - name: sources
      dtype: string
    - name: references
      dtype: string
    - name: language_pair
      dtype: string
    - name: dataset
      dtype: string
  splits:
    - name: train
      num_bytes: 14008918721
      num_examples: 241828
  download_size: 7996471024
  dataset_size: 14008918721
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
license: cc-by-sa-4.0
task_categories:
  - translation
language:
  - en
  - de
  - es
  - fr
  - it
  - ko
  - nl
  - pt
  - ru
  - zh
size_categories:
  - 100K<n<1M

Dataset Card for DocBlocks

DocBlocks is a high-quality, multilingual document-level machine translation (MT) dataset designed to fine-tune large language models (LLMs) on long-context translation tasks. Unlike traditional sentence-level datasets, it contains full documents with natural discourse structures and contextual alignment, helping models maintain coherence, consistency, and high translation quality across longer texts.

Curated by: Instituto Superior Técnico, Instituto de Telecomunicações, Carnegie Mellon University and Unbabel;
Language(s) (NLP): English, German, Spanish, French, Italian, Dutch, Portuguese, Russian, Korean, Chinese;
License: DocBlocks includes data from the following sources: IWSLT, Europarl, News Commentary, GuoFeng, and BWB. For licensing information, please refer to the official documentation or websites of each source.

Dataset Details

conversations - The user and assistant dialogue turns, following an instruction-based format suitable for LLM fine-tuning;
sources - The original source text in the source language;
references - The human-translated target/reference text in the target language;
language_pair - The language direction of the translation pair;
dataset - The name of the original dataset from which the document was sourced.

Bias, Risks, and Limitations

DocBlocks may reflect linguistic, cultural, and domain biases from its source corpora, and its performance is influenced by language coverage and document structure variability.

Citation

@misc{multilingual_contextualization_llm_2025,
      title={Multilingual Contextualization of Large Language Models for Document-Level Machine Translation}, 
      author={Miguel Moura Ramos and Patrick Fernandes and Sweta Agrawal and André F. T. Martins},
      year={2025},
      eprint={2504.12140},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2504.12140}, 
}