TildeOpen-MT-Euro-mix

TildeOpen-MT-Euro-mix is a machine-translation datamix for 30 European languages. It was built for the translation instruction-tuning experiments used to validate the TildeOpen LLM (Bergmanis et al., LREC 2026). We publish this data to support the reproducibility of our experiments.

Composition

Datamix contains 1,890,652 sentence/document pairs across 245 translation directions and 30 languages - 221.44M source-side and 240.56M target-side (trainable) tokens. Token counts are computed with the TildeAI/TildeOpen-30b-64k tokeniser, without special tokens.

It combines authentic and synthetic parallel data from four sources:

• Eur-Lex: recrawled European Union legal texts.
• TED: segments of TED talk transcript translations as distributed by OPUS.
• OpenSubtitles: segments of film and TV subtitle translations as distributed by OPUS.
• backtranslated-news: synthetic pairs built from WMT News Crawl monolingual news: authentic target-side text paired with Gemma 3 27b machine-translated English on the source side.

The mix was subsampled to obtain target-token-side-balanced corpora. Every record carries provenance metadata, so authentic and synthetic pairs can be separated at load time.

Origin	Records	Directions	Source tokens	Target tokens
backtranslated-news	273,308	29	141,258,774	159,040,806
OpenSubtitles	1,503,440	224	57,117,564	58,382,402
TED	95,859	209	11,495,752	11,604,204
eur-lex	18,045	131	11,564,314	11,534,755

Usage

from datasets import load_dataset

ds = load_dataset("TildeAI/TildeOpen-MT-Euro-mix", split="train")

# e.g. keep only pairs with no machine-generated side:
authentic = ds.filter(lambda r: r["synthetic_side"] == "")

Each record is directional: build training examples from source to target, using source_lang/target_lang for the direction.

Citation

If you use this data mixture, please cite the TildeOpen paper:

@inproceedings{bergmanis-etal-2026-tildeopen,
  title     = {{TildeOpen} {LLM}: Leveraging Curriculum Learning to Achieve
               Equitable Language Representation},
  author    = {Bergmanis, Toms and Kronis, Martins and
               Pretkalni{\c{n}}{\v{s}}, Ingus J{\=a}nis and
               Nicmanis, Davis and Je{\c{l}}inska, Je{\c{l}}izaveta and
               Rozis, Roberts and V{\=\i}ksna, Rinalds and
               Pinnis, M{\=a}rcis},
  booktitle = {Proceedings of the Fifteenth Language Resources and
               Evaluation Conference ({LREC} 2026)},
  pages     = {10634--10652},
  year      = {2026},
  publisher = {ELRA Language Resources Association}
}

Acknowledgements

Parallel segments for TED and OpenSubtitles are used as distributed by the OPUS collection; monolingual news comes from the WMT News Crawl; EU legal texts originate from EUR-Lex. Please also acknowledge these upstream resources:

@inproceedings{tiedemann-2012-parallel,
  title     = {Parallel Data, Tools and Interfaces in {OPUS}},
  author    = {Tiedemann, J{\"o}rg},
  booktitle = {Proceedings of the Eighth International Conference on Language
               Resources and Evaluation ({LREC} 2012)},
  year      = {2012}
}

@inproceedings{lison-tiedemann-2016-opensubtitles,
  title     = {{OpenSubtitles2016}: Extracting Large Parallel Corpora from
               Movie and {TV} Subtitles},
  author    = {Lison, Pierre and Tiedemann, J{\"o}rg},
  booktitle = {Proceedings of the Tenth International Conference on Language
               Resources and Evaluation ({LREC} 2016)},
  year      = {2016}
}

@inproceedings{salesky-etal-2021-tedx,
  title     = {The Multilingual {TEDx} Corpus for Speech Recognition and
               Translation},
  author    = {Salesky, Elizabeth and Wiesner, Matthew and Bremerman, Jacob
               and Cattoni, Roldano and Negri, Matteo and Turchi, Marco and
               Oard, Douglas W. and Post, Matt},
  booktitle = {Proceedings of Interspeech 2021},
  year      = {2021}
}

Disclaimer and takedown

Tilde does not own the text from which this dataset was extracted. The underlying content belongs to its original authors and rightsholders; this dataset redistributes only segments that upstream distributors (OPUS, WMT, EUR-Lex) make publicly available, together with machine-generated translations of such content, and claims no rights over any of it.

If you believe that any content in this dataset infringes your rights, please contact us by opening a discussion on this repository or by writing to tilde@tilde.lv, and we will review the request and remove the material in question.