SiLTT dataset

SiLTT (Silesian Literary Translation Testset) is a Polish–Silesian parallel test set derived from human translations and automatically aligned using LaBSE sentence embeddings.

The dataset was compiled from literary and cultural texts, including Cinderella, Rapunzel, Janko the Musician, and A Christmas Carol. The original version contains 1,237 sentence pairs. We also provide SiLTT_filtered, a lightly revised version with 1,219 pairs.

⚠️ Note: This version may still contain some misaligned segments. We’re continuing to refine the filtering and add new examples.

Statistic	Polish	Silesian
Avg. sentence length (chars)	93.1	79.4
Avg. sentence length (words)	14.8	13.6
Max. sentence length (chars)	542	445
Max. sentence length (words)	82	73
Vocabulary size	7,673	5,393

Usage

from datasets import load_dataset

dataset = load_dataset("NASK-PIB/SiLTT")
print(dataset)

Output:

DatasetDict({
    SiLTT_filtered: Dataset({
        features: ['pol', 'szl'],
        num_rows: 1219
    })
    SiLTT: Dataset({
        features: ['pol', 'szl'],
        num_rows: 1237
    })
})

License

SiLTT is released under the CC BY-NC 4.0 license.

Authors

• Grzegorz Kulik — grzegorz.kulik@nask.pl
• Mikołaj Pokrywka — mikolaj.pokrywka@nask.pl
• Adam Jatowt
• Wojciech Kusa — wojciech.kusa@nask.pl

Cite

@misc{kulik2026precisionscalepolishsilesianbenchmark,
      title={Precision over Scale: A Polish-Silesian Benchmark and a Translation System Outperforming Open-Source and Commercial Models}, 
      author={Grzegorz Kulik and Mikołaj Pokrywka and Adam Jatowt and Wojciech Kusa},
      year={2026},
      eprint={2610.01082},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2610.01082}, 
}