lightonai/multilingual-embeddings-pre-training-curated

📚 Collection | 📝 Multilingual Blog | 📝 English Blog

Contrastive Multilingual Pre-Training

2.16B query–document pairs across eight languages, plus cross-lingual pairs

mDenseOn | mLateOn | DenseOn | LateOn | PyLate | FastPlaid

🎯 TL;DR: The multilingual contrastive pre-training corpus used to train mDenseOn and mLateOn. It extends our curated English data recipe (embeddings-pre-training-curated) to French, German, Italian, Spanish, Portuguese, Swedish, Norwegian, and Arabic through translate-train, and adds a cross-lingual split where queries and documents are in different languages. Together with the 665M-pair English seed, it forms the ~2.8B-pair corpus behind our multilingual retrievers — one of the largest open multilingual retrieval training sets to date.

Overview

High-quality retrieval supervision is overwhelmingly concentrated in English. Rather than collecting multilingual corpora from scratch — expensive, uneven across languages, and hard to curate to a consistent quality — we applied the translate-train strategy: machine-translating our validated English pre-training data into eight target languages, and deriving cross-lingual pairs from the aligned translations.

Each English query–document pair was translated jointly (query and document in the same prompt, so both stay grounded in the same context) using Mistral-Small-3.1-24B-Instruct, with language-specific prompts and two in-language examples encouraging idiomatic reformulation rather than literal transliteration.

Because translation starts from the already-curated English seed, this corpus directly inherits its filtering: 30+ source-aware structural filters, MD5 deduplication, and cross-encoder relevance scoring with mxbai-rerank-large-v2 (see the curated English card for the full pipeline). On top of the translated subsets, each language also carries organic in-language data (FineWeb2 and Wikipedia) drawn from the parent corpus.

A key finding from training on this data: with a late-interaction model, translate-train into a handful of representative languages is enough for broad multilingual generalization — mLateOn transfers to languages and scripts never seen during retrieval training (Cyrillic, CJK, Devanagari, Thai, and more). See the multilingual blog post and paper for the full analysis.

The DenseOn / LateOn Data Family

Dataset	Description	Scale
embeddings-pre-training	Raw annotated English parent corpus (non-destructive filter flags)	1.4B pairs
embeddings-pre-training-curated	Curated English seed — the exact English training slice	665M pairs
contrastive-multilingual (this card)	Translate-train extension to 8 languages + cross-lingual pairs	2.16B pairs
embeddings-fine-tuning	English fine-tuning data with 2,048 mined negatives per query	1.88M examples
embeddings-fine-tuning-filtered-{en,fr,es,it,de,pt,ar,no,sv}	Multilingual fine-tuning data (one repository per language)	~1.56M examples each
embeddings-fine-tuning-filtered-code	Code fine-tuning data covering the MTEB Code training splits	2.12M examples

Dataset Structure

The repository is organized as one folder per target language, plus one folder of cross-lingual pairs:

contrastive-multilingual/
├── ar/            # Arabic      — 236M pairs
├── de/            # German      — 243M pairs
├── es/            # Spanish     — 242M pairs
├── fr/            # French      — 354M pairs
├── it/            # Italian     — 240M pairs
├── no/            # Norwegian   — 224M pairs
├── pt/            # Portuguese  — 166M pairs
├── sv/            # Swedish     — 238M pairs
└── cross-lingual/ # mixed       — 220M pairs

Inside each language folder, files keep the same subset names as the English seed (msmarco-*, reddit-*, paq-*, …): the subset name identifies the original source, the folder identifies the language of the text.

Each row has three columns:

Column	Type	Description
index	int64	Stable row identifier, shared with the English parent corpus and across all languages
query	string	The query text
document	string	The paired document text

Aligned by design: the index column

Because every language is translated from the same English seed, index acts as a join key across the whole family: the same index within the same subset refers to the same underlying pair in every language folder and in embeddings-pre-training-curated. For example, index = 76081 of wikihow:

Language	Query
English (seed)	Memorize an Essay
French	Comment mémoriser un essai ?
Italian	Come memorizzare un saggio?
Arabic	كيفية حفظ مقالة

This makes it easy to build your own parallel corpora, cross-lingual pairs, or per-language ablations without re-translating anything.

The cross-lingual split

To support cross-lingual retrieval, a subset of pairs is materialized with the query in one language and the document in a uniformly sampled different language (English included):

Query	Document
Är Narendra Modi överhyllad i Indien? (Swedish)	Narendra Modi est-il vraiment surévalué ? Pourquoi ? (French)
ما السبب الذي يجعلني لم أجد صديق حتى الآن؟ (Arabic)	What could be the reason I never had a boyfriend? (English)

Per-language statistics

Config	Language	Pairs	Files	Size
ar	Arabic	235.7M	783	125.5 GB
de	German	243.2M	797	130.0 GB
es	Spanish	242.2M	794	123.4 GB
fr	French	354.0M	1,159	197.5 GB
it	Italian	239.7M	793	119.7 GB
no	Norwegian	224.0M	745	101.0 GB
pt	Portuguese	165.6M	569	84.7 GB
sv	Swedish	237.8M	785	108.5 GB
cross-lingual	mixed	220.1M	781	119.8 GB
Total		2.16B	7,206	~1.1 TB

Notes:

• Translated subsets (the vast majority) are joint machine translations of the curated English seed. Subsets named after non-English sources (cc_news_fr, webfaq_fra, webfaq_deu, …) are likewise translations of those sources into the folder's language.
• Organic subsets — fw2-* (FineWeb2) and wikipedia_* — are native in-language data from the parent corpus, not translations. Norwegian has no FineWeb2 subset (FineWeb2 does not cover it).
• French is larger because it additionally contains mtp (113.1M pairs) and reddit_body_comment (7.5M pairs), which were only translated into French.

Full subset × language matrix (pairs per subset)

Subset	ar	de	es	fr	it	no	pt	sv	cross‑lingual
agnews	0.4M	0.4M	0.4M	0.4M	0.4M	0.4M	0.4M	0.4M	0.4M
altlex	62k	62k	62k	62k	62k	62k	—	62k	55k
amazon_qa	0.4M	0.4M	0.4M	0.4M	0.4M	0.4M	0.4M	0.4M	0.4M
amazon_reviews	25.1M	25.1M	25.1M	25.1M	25.1M	24.4M	19.0M	25.1M	24.3M
arxiv_title_abstract	1.4M	1.4M	1.4M	1.4M	1.4M	1.4M	1.1M	1.4M	1.4M
beir_dbpedia	1.6M	1.6M	1.6M	1.6M	1.6M	1.2M	0.4M	1.6M	1.4M
biorxiv_title_abstract	0.2M	0.2M	0.2M	0.2M	0.2M	0.2M	0.2M	0.2M	0.2M
cc_news_fr	9.8M	9.8M	10.1M	11.2M	10.1M	9.5M	7.8M	9.9M	8.3M
ccnews_en	0.2M	0.2M	0.2M	0.2M	0.2M	0.2M	0.1M	0.2M	0.2M
cnn_dailymail	0.2M	0.2M	0.2M	0.2M	0.2M	0.2M	0.1M	0.2M	0.2M
codesearchnet	0.3M	0.3M	0.3M	0.7M	0.7M	0.3M	0.3M	0.7M	0.5M
eli5	0.2M	0.2M	0.2M	0.2M	0.2M	0.2M	0.2M	0.2M	0.2M
fw2-arb_Arab	0.8M	—	—	—	—	—	—	—	—
fw2-deu_Latn	—	6.9M	—	—	—	—	—	—	—
fw2-fra_Latn	—	—	—	5.1M	—	—	—	—	—
fw2-ita_Latn	—	—	—	—	3.0M	—	—	—	—
fw2-por_Latn	—	—	—	—	—	—	2.9M	—	—
fw2-spa_Latn	—	—	6.2M	—	—	—	—	—	—
fw2-swe_Latn	—	—	—	—	—	—	—	0.6M	—
gooaq_qa	1.9M	1.9M	1.9M	1.9M	1.9M	1.9M	1.9M	1.9M	1.9M
hermes	0.6M	0.6M	0.6M	0.4M	0.6M	0.6M	0.6M	0.6M	0.6M
medrxiv_title_abstract	0.1M	0.1M	0.1M	0.1M	0.1M	0.1M	—	0.1M	0.1M
msmarco	2.3M	2.3M	2.3M	2.3M	2.3M	2.3M	2.3M	2.3M	2.3M
mtp	—	—	—	113.1M	—	—	—	—	—
npr	0.3M	0.3M	0.3M	0.3M	0.3M	0.3M	0.3M	0.3M	0.3M
paq	36.3M	36.3M	36.2M	36.3M	36.3M	34.8M	22.7M	36.3M	34.4M
quora	34k	34k	34k	34k	34k	34k	34k	34k	34k
reddit	50.4M	50.7M	50.7M	53.7M	50.7M	47.4M	37.5M	50.7M	48.7M
reddit_body_comment	—	—	—	7.5M	—	—	—	—	—
s2orc_abstract_citation	8.8M	8.8M	8.9M	8.8M	8.9M	8.6M	6.7M	8.9M	8.5M
s2orc_citation_titles	7.5M	7.9M	7.9M	—	7.9M	6.3M	4.3M	7.9M	6.4M
s2orc_title_abstract	25.8M	25.9M	25.9M	25.9M	25.9M	25.1M	19.3M	25.9M	25.0M
sentence_compression	0.1M	0.1M	0.1M	0.1M	0.1M	0.1M	0.1M	0.1M	0.1M
simplewiki	73k	73k	73k	73k	73k	73k	—	73k	65k
stackexchange_body_body	41k	41k	41k	41k	41k	41k	—	41k	36k
stackexchange_duplicate_questions	40k	40k	40k	40k	40k	40k	40k	40k	40k
stackexchange_qa	0.9M	0.3M	0.7M	2.2M	0.9M	0.9M	0.6M	0.9M	0.8M
stackexchange_title_body	3.3M	3.3M	3.3M	3.3M	3.3M	3.0M	2.3M	3.3M	3.2M
stackoverflow_title_body	9.4M	9.4M	9.3M	9.2M	9.4M	9.2M	6.5M	9.4M	8.8M
webfaq_deu	4.7M	4.7M	4.7M	4.7M	4.7M	4.7M	1.8M	4.7M	3.8M
webfaq_eng	31.5M	31.5M	31.5M	25.4M	31.5M	30.5M	20.1M	31.5M	29.4M
webfaq_fra	2.9M	2.9M	2.9M	2.9M	2.9M	2.9M	2.2M	2.9M	2.5M
webfaq_ita	1.8M	1.8M	1.8M	1.8M	1.8M	0.9M	—	1.8M	1.3M
webfaq_spa	4.0M	4.0M	4.0M	4.0M	4.0M	4.0M	1.6M	4.0M	3.3M
wikihow	88k	88k	88k	88k	88k	88k	88k	88k	88k
wikipedia_ar	0.9M	—	—	—	—	—	—	—	—
wikipedia_de	—	2.1M	—	—	—	—	—	—	—
wikipedia_es	—	—	1.4M	—	—	—	—	—	—
wikipedia_fr	—	—	—	1.9M	—	—	—	—	—
wikipedia_it	—	—	—	—	1.2M	—	—	—	—
wikipedia_no	—	—	—	—	—	0.5M	—	—	—
wikipedia_pt	—	—	—	—	—	—	0.8M	—	—
wikipedia_sv	—	—	—	—	—	—	—	1.9M	—
yahoo_answer	0.3M	0.3M	0.3M	0.3M	0.3M	0.3M	0.3M	0.3M	0.3M
yahoo_qa	0.5M	0.5M	0.5M	0.5M	0.5M	0.5M	0.2M	0.5M	0.4M
yahoo_question_body	0.2M	0.2M	0.2M	0.2M	0.2M	0.2M	—	0.2M	0.2M
Total	236M	243M	242M	354M	240M	224M	166M	238M	220M

For per-source documentation and license pointers, see the curated English dataset card — every subset here maps to a source documented there.

Quick Start

from datasets import load_dataset

# One language (streaming recommended — each config is 85–200 GB)
dataset = load_dataset(
    "lightonai/contrastive-multilingual",
    "fr",
    split="train",
    streaming=True,
)

# A single subset of a single language
dataset = load_dataset(
    "lightonai/contrastive-multilingual",
    data_files="fr/msmarco-*.parquet",
    split="train",
)

# Cross-lingual pairs (query and document in different languages)
dataset = load_dataset(
    "lightonai/contrastive-multilingual",
    "cross-lingual",
    split="train",
    streaming=True,
)

Build your own parallel data by joining languages on index:

fr = load_dataset("lightonai/contrastive-multilingual", data_files="fr/wikihow-*.parquet", split="train").to_pandas()
it = load_dataset("lightonai/contrastive-multilingual", data_files="it/wikihow-*.parquet", split="train").to_pandas()

parallel = fr.merge(it, on="index", suffixes=("_fr", "_it"))

How We Trained on It

For mDenseOn and mLateOn, this corpus was mixed with the curated English seed using the following recipe (full details and ablations in the blog post):

• Mixture: 20% English (embeddings-pre-training-curated) / 80% multilingual, with a 25% cross-lingual share inside the multilingual portion.
• Sampling: multinomial language/source sampling with α = 0.5, so large sources do not dominate.
• Batching: each batch is drawn from a single source, preventing the model from exploiting source-specific shortcuts; batch size 16,384 with in-batch negatives (InfoNCE), using GradCache to keep memory manageable.

Interestingly, our mixture ablations showed that final performance is fairly robust to the English/multilingual ratio, and that a modest amount of cross-lingual data suffices — translate-train already provides a strong alignment signal on top of the backbone's multilingual pre-training.

Models trained on this dataset

Both models are 307M parameters, built on the mmBERT-base backbone (NDCG@10; tgt = benchmark languages overlapping the training languages):

Model	BEIR	MIRACLtgt	MIRACL	MLDRtgt	MLDR	MTEB Code
mLateOn	57.56	65.61	67.04	87.69	77.92	73.48
mDenseOn	56.70	59.61	58.02	64.98	51.59	71.53

Although this corpus covers eight target languages, mLateOn generalizes far beyond them — it is evaluated on 18 MIRACL languages while only five overlap with training, and scores higher on the full benchmark than on the target subset. Late interaction turns translate-train from a target-language expansion strategy into a multilingual generalization recipe.

Licensing

This dataset is not openly licensed. It is derived by machine translation from sources that each retain their original license; we do not relicense any data. Users are responsible for verifying that their intended use complies with the license terms of each individual source. The curated English dataset card links every subset to its original source, where license information can be found.

Limitations

The corpus is predominantly synthetic: most pairs are machine translations of English supervision. Translation can introduce artifacts, preserve English-specific assumptions, flatten culturally specific distinctions, and differ from naturally occurring queries and documents in the target languages. Joint translation of query and document may also make some positive pairs easier than organic retrieval examples, since both sides inherit lexical and stylistic choices from the same source. The organic FineWeb2 and Wikipedia subsets partially mitigate this, but users should not treat this corpus as a substitute for natively collected data in every application.

Citation

If you are using this dataset, please consider citing our work

@misc{sourty2026denseonlateonfullyopen,
  title         = {DenseOn with the LateOn: Fully Open Dense and Late-Interaction Models for Multilingual, Long-Context, and Code Search},
  author        = {Raphaël Sourty and Antoine Chaffin and Paulo Roberto Moura Junior and Amélie Chatelain},
  year          = {2026},
  eprint        = {2607.27178},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2607.27178},
}

Acknowledgements

This work was granted access to the HPC resources of IDRIS under GENCI allocations AS011016449, A0181016214, and A0171015706 (Jean Zay supercomputer). We also acknowledge the Barcelona Supercomputing Center (BSC-CNS) for providing access to MareNostrum 5 under EuroHPC AI Factory Fast Lane project EHPC-AIF-2025FL01-445. This project is also supported by the OpenEuroLLM project, co-funded by the Digital Europe Programme under GA no. 101195233.