AdleBenSalem/flores-plus-doc-level

FLORES+ Document-Level (en / de / es / fr / ru)

A document-level regrouping of FLORES+ for English, German, Spanish, French and Russian. FLORES+ sentences were originally extracted as small clusters of consecutive sentences from the same web article; this dataset restores that structure so the benchmark can be used for paragraph- and document-level MT evaluation and multilingual retrieval, not just isolated sentences.

What's in it

Two configs, both multi-parallel across the five languages (text_en, text_de, text_es, text_fr, text_ru):

config	rows (dev / devtest)	one row is
documents (default)	281 / 281	one source article: all its sentences concatenated in order
sentences	997 / 1012	one sentence, with doc_id + position for grouping

Documents contain 3.5 sentences on average (max 10 in dev, 6 in devtest). Fields common to both configs: url (source article), domain, topic. documents additionally has n_sentences and sentence_ids; sentences keeps the original FLORES+ id so rows can be joined back to the upstream dataset and to any other language.

How it was built

1. Load the five languoid configs (eng_Latn, deu_Latn, spa_Latn, fra_Latn, rus_Cyrl) of openlanguagedata/flores_plus for dev and devtest, and keep the rows whose id is present in all languages (row-aligned multi-parallel corpus; nothing was dropped in practice).
2. Sort by id and group consecutive runs of the same url into documents — FLORES articles are contiguous in id-order, so this restores the original sentence clusters exactly.
3. documents rows concatenate the sentences of one article with single spaces, per language.

Construction code: Jeerhz/FDTEM.

Block inventory (what fits in a block of k segments)

A common way to use this dataset is to segment each document into blocks of k consecutive segments (non-overlapping, i.e. stride k; tails shorter than k are dropped). This is how many blocks that yields:

k (segments per block)	dev	devtest	combined
1	997	1012	2009
2	417	429	846
3	234	234	468
4	130	146	276
5	70	71	141
6	5	3	8

Document sizes behind these counts — dev: 53 docs of 2 sentences, 99 of 3, 60 of 4, 64 of 5, 4 of 6, 1 of 10; devtest: 1 doc of 1 sentence, 49 of 2, 85 of 3, 75 of 4, 68 of 5, 3 of 6.

⚠️ Fairness disclaimer. Long blocks are rare: the number of available blocks falls sharply with k (2009 → 846 → 468 → 276 → 141 → 8), and blocks with many segments can only come from the minority of long articles. Experiments that compare results across k should account for this — both statistically (smaller samples at large k mean wider confidence intervals; report uncertainty, e.g. via bootstrap) and compositionally (large-k results are dominated by a few long documents, so per-k populations are not identically distributed). k ≥ 6 is effectively anecdotal in this corpus and should not carry conclusions.

Intended use

Sentences that share a document form natural, variable-length contexts: the two configs make it easy to build inputs of controlled length out of real, coherent text, aligned across five languages. We use it to study how multilingual encoders and MT-evaluation models behave as inputs grow beyond the single sentence — retrieval, evaluation and probing tasks at the block and document level. Accompanying code lives in Jeerhz/FDTEM.

Licensing & attribution

Derivative of FLORES+ by the Open Language Data Initiative (sentences originally translated for FLORES-101/200 by Meta AI), released — like its source — under CC-BY-SA-4.0. No sentence text was modified; only the grouping/concatenation structure was added.

Citation

If you use this dataset, cite FLORES+ / FLORES-200:

@inproceedings{dale-costa-jussa-2024-flores,
  title     = "{FLORES}+: an extended benchmark and toolkit for multilingual machine translation evaluation",
  author    = "Dale, David and Costa-juss{\`a}, Marta R.",
  booktitle = "Proceedings of the Ninth Conference on Machine Translation (WMT)",
  year      = "2024"
}
@article{nllb2022,
  title  = "No Language Left Behind: Scaling Human-Centered Machine Translation",
  author = "{NLLB Team} and others",
  journal = "arXiv:2207.04672",
  year   = "2022"
}