DocHPLT: A Massively Multilingual Document-Level Translation Dataset

Existing document-level machine translation resources are only available for a handful of languages, mostly high-resourced ones. To facilitate the training and evaluation of document-level translation and, more broadly, long-context modeling for global communities, we create DocHPLT, the largest publicly available document-level translation dataset to date. It contains 124 million aligned document pairs across 50 languages paired with English, comprising 4.26 billion sentences, with further possibility to provide 2500 bonus pairs not involving English. Unlike previous reconstruction-based approaches that piece together documents from sentence-level data, we modify an existing web extraction pipeline to preserve complete document integrity from the source, retaining all content including un- aligned portions. After our preliminary experiments identify the optimal training context strategy for document-level translation, we demonstrate that LLMs fine-tuned on DocHPLT substantially outperform off-the-shelf instruction- tuned baselines, with particularly dramatic improvements for under-resourced languages. We open-source the dataset under a permissive license, providing essential infrastructure for advancing multilingual document-level translation.

Corpus statistics

	#sentences	#docs
af	16,416,841	297,636
ar	65,482,300	2,271,167
az	12,202,189	332,742
be	10,672,952	212,121
bg	80,018,549	1,746,301
bn	10,473,372	414,099
bs	20,635,243	514,615
ca	47,905,003	1,198,217
cy	8,908,119	265,261
en	2,665,945,834	47,484,349
eo	6,115,355	119,196
et	33,684,509	774,561
eu	6,783,654	189,347
fa	24,837,952	810,029
fi	111,615,913	2,445,791
ga	6,398,081	172,167
gl	10,657,570	233,545
gu	3,202,679	108,507
he	38,077,820	1,190,198
hi	37,592,475	1,336,090
hr	52,267,826	1,063,347
is	12,571,982	274,078
ja	164,136,152	4,032,689
kk	5,948,866	140,082
kn	4,463,262	123,053
ko	84,527,642	2,058,811
lt	48,692,264	1,031,628
lv	37,426,957	796,659
mk	12,465,228	307,055
ml	2,925,457	115,189
mr	3,066,703	128,808
ms	51,150,528	978,185
mt	6,328,544	141,088
nb	89,189,502	1,884,362
ne	1,549,852	74,579
nn	4,228,079	93,285
si	1,497,375	50,605
sk	70,057,465	1,461,804
sl	37,501,647	797,858
sq	11,475,561	328,651
sr	21,620,629	407,440
sw	8,409,824	185,287
ta	6,790,864	215,564
te	5,131,680	141,279
th	16,134,265	676,699
tr	100,380,235	3,884,137
uk	89,841,883	1,955,041
ur	5,479,098	234,708
uz	3,502,356	69,440
vi	87,511,126	1,986,258
xh	995,556	21,561
total	4,264,894,818	87,775,169

Link for arXiv preprint: https://arxiv.org/abs/2508.13079

This project has received funding from the European Union's Horizon Europe research and innovation programme under grant agreement No 101070350 and from UK Research and Innovation (UKRI) under the UK government’s Horizon Europe funding guarantee [grant number 10052546].

Citation

If you use this resource, please kindly cite:

@article{dochplt,
      title={{DocHPLT}: A Massively Multilingual Document-Level Translation Dataset}, 
      author={Dayyán O'Brien and Bhavitvya Malik and Ona de Gibert and Pinzhen Chen and Barry Haddow and Jörg Tiedemann},
      year={2025},
      journal={arXiv preprint},
      url={https://arxiv.org/abs/2508.13079}, 
}