lightonai/embeddings-pre-training-curated

Embeddings pre-training curated data

This dataset is the English subset of lightonai/embeddings-pre-training, assembled to reproduce the English data recipe described in the mGTE technical report (Zhang et al., 2024).

The mGTE paper describes the data sources used to train the GTE family of multilingual text embedding and reranking models, but does not release the data itself. This dataset is our reconstruction of the English portion of that recipe, curated as part of a research effort to understand how data composition affects retrieval model quality. For more information please check our blogpost.

For the full multilingual collection (50+ subsets across multiple languages), see the parent dataset: lightonai/embeddings-pre-training

---

Licensing

This dataset is not openly licensed. Each source retains its 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 before downloading or using this data. The "Original Source" column in the tables below links to where license information can be found.

---

Dataset Structure

Each row is a text pair with the following columns:

Column	Type	Description
index	int64	Row identifier, inherited from the parent embeddings-pre-training dataset
query	string	The input text
document	string	The corresponding document text
similarity	float32	Query–document relevance score from a cross-encoder reranker (mxbai-rerank-large-v2)

Note on schema vs parent dataset. The parent lightonai/embeddings-pre-training additionally carries drop (bool) and duplicate (int64) columns produced by the per-source filter pipeline and MD5 deduplication. Those annotations have already been applied to produce embeddings-pre-training-curated (see "Curation & Filtering" below), so they are no longer present in the shipped parquet files. If you need access to the unfiltered pool with the raw annotations, go back to the parent dataset.

---

Quick Start

from datasets import load_dataset

# Load a specific subset
dataset = load_dataset(
    "lightonai/embeddings-pre-training-curated",
    "msmarco",
    split="train",
)

Every row in embeddings-pre-training-curated has already passed the recommended curation pipeline, so no post-filter is required. If you want to be stricter on semantic relevance, raise the similarity floor:

dataset = dataset.filter(lambda x: x["similarity"] >= 5.0)

---

Curation & Filtering

embeddings-pre-training-curated is derived from lightonai/embeddings-pre-training by applying, in this order:

1. Source-aware rule-based filters — a per-source pipeline of up to 18 filters (policy boilerplate, HTML artifacts, bad control chars, non-target scripts, language identification via FastText, uppercase / numeric ratios, Google 1T unigram log-probability, repeated-uncommon-word, minimum token count, strict allow-listed character set). Rows are annotated with a boolean drop flag; all rows with drop = True are removed here.

2. MD5 deduplication — an MD5 hash of query + " " + document marks every row beyond the first occurrence of its hash with the duplicate index column pointing at the canonical row. All rows with duplicate IS NOT NULL are removed here.

3. Cross-encoder relevance scoring — every remaining query–document pair is scored by mxbai-rerank-large-v2 into the similarity column.

4. Similarity threshold — for every subset except fw_edu we keep pairs with similarity >= 3.0.

5. Self-pair removal — the residual case query == document (identical strings) is removed. This catches self-pair rows that MD5 dedup does not flag (dedup only finds cross-row collisions, not rows whose query equals its own document).

The SQL-equivalent filter applied to every standard subset is:

SELECT index, query, document, similarity
FROM lightonai/embeddings-pre-training
WHERE NOT drop
  AND duplicate IS NULL
  AND similarity >= 3.0
  AND query <> document

Per-source retention (sampled)

Retention ratios measured on one parquet shard per subset:

Subset	Raw rows	Kept	Retention
agnews	1 157 745	564 258	48.7 %
altlex	110 708	83 053	75.0 %
amazon_qa	1 095 290	761 984	69.6 %
biorxiv_title_abstract	283 550	275 247	97.1 %
arxiv_title_abstract (shard 0/5)	399 898	372 315	93.1 %

The wide retention spread reflects intrinsic source quality rather than filter aggressiveness: curated scientific abstracts lose almost nothing, while noisier web-crawled news lose ~half.

Special case: fw_edu (FineWeb-Edu)

Because fw_edu is produced by an upstream retrieval-common-crawl pipeline (see orionweller/contrastive-pretraining) that already cleans and deduplicates at the page level, applying our surface-rule filter or a second MD5 dedup would be both redundant and prohibitively expensive at ~400 M rows. For fw_edu only:

• No rule-based filter applied (drop = False on every row).
• No MD5 dedup applied (duplicate = NULL on every row).
• Cross-encoder-only curation. Instead of the similarity >= 3.0 floor used on every other subset, we keep the top ~34 % of pairs per shard by cross-encoder similarity. The effective absolute similarity floor varies between ~10.6 and ~11.1 across shards (versus 3.0 for every other subset).

Special case: the Atlas HLP Wikipedia splits

The two subsets wikipedia_hlp_cm and wikipedia_hlp_dl (10 M rows each, from facebookresearch/atlas) are passed through untouched: no rule-based filter, no MD5 dedup, and no cross-encoder scoring (their similarity column is a placeholder zero, preserved only for schema consistency). These are the Atlas paragraph-linking pairs as published.

---

Subsets

34 subsets | 1 235 files | ~517 GB total

News & Media

Subset	Files	Size	Original Source
agnews	1	0.10 GB	sentence-transformers/agnews
cc_news_en	2	0.41 GB	nomic-ai/nomic-embed-unsupervised-data
cnn_dailymail	3	0.68 GB	sentence-transformers/embedding-training-data
npr	3	0.53 GB	sentence-transformers/npr

Scientific & Academic

Subset	Files	Size	Original Source
arxiv_title_abstract	5	1.11 GB	UniverseTBD/arxiv-abstracts-large
biorxiv_title_abstract	1	0.26 GB	laion/biorXiv_metadata
medrxiv_title_abstract	1	0.18 GB	mteb/raw_medrxiv
s2orc_abstract_citation	185	34.36 GB	sentence-transformers/s2orc
s2orc_citation_titles	20	3.47 GB	sentence-transformers/s2orc
s2orc_title_abstract	63	15.94 GB	sentence-transformers/s2orc

QA & Information Retrieval

Subset	Files	Size	Original Source
amazon_qa	1	0.15 GB	nomic-ai/nomic-embed-unsupervised-data
gooaq_qa	2	0.50 GB	sentence-transformers/embedding-training-data
msmarco	3	0.91 GB	microsoft/ms_marco
paq	75	21.95 GB	sentence-transformers/paq
quora	1	< 0.01 GB	nomic-ai/nomic-embed-unsupervised-data
yahoo_answer	1	0.27 GB	sentence-transformers/embedding-training-data
yahoo_qa	2	0.28 GB	sentence-transformers/embedding-training-data
yahoo_question_body	1	0.10 GB	sentence-transformers/embedding-training-data

Reviews & Commerce

Subset	Files	Size	Original Source
amazon_reviews	33	8.59 GB	sentence-transformers/amazon-reviews

Social & Forum

Subset	Files	Size	Original Source
reddit	188	36.86 GB	sentence-transformers/reddit
reddit_body_comment	45	11.84 GB	HuggingFaceGECLM/REDDIT_submissions
stackexchange_body_body	1	0.04 GB	sentence-transformers/embedding-training-data
stackexchange_duplicate_questions	1	0.01 GB	sentence-transformers/embedding-training-data
stackexchange_qa	10	2.18 GB	sentence-transformers/embedding-training-data
stackexchange_title_body	10	2.24 GB	sentence-transformers/embedding-training-data
stackoverflow_title_body	42	7.54 GB	sentence-transformers/embedding-training-data

Encyclopedia & Reference

Subset	Files	Size	Original Source
beir_dbpedia	4	0.49 GB	BeIR/dbpedia-entity
wikianswers	41	0.75 GB	sentence-transformers/embedding-training-data
wikihow	1	0.02 GB	sentence-transformers/embedding-training-data
wikipedia_hlp_cm	1	4.73 GB	facebookresearch/atlas
wikipedia_hlp_dl	1	4.81 GB	facebookresearch/atlas

NLP & Paraphrase

Subset	Files	Size	Original Source
altlex	1	0.02 GB	sentence-transformers/altlex
mtp	367	98.26 GB	mGTE paper (Massive Text Pairs)

Web & Education

Subset	Files	Size	Original Source
fw_edu	119	257.04 GB	orionweller/contrastive-pretraining (derived from HuggingFaceFW/fineweb-edu)

---

Citation

If you use this dataset, please cite both works:

@misc{sourty2025denseonlateon,
  title={DenseOn with LateOn: Open State-of-the-Art Single and Multi-Vector Models},
  author={Sourty, Raphael and Chaffin, Antoine and Weller, Orion and Demoura, Paulo and Chatelain, Amélie},
  year={2026},
  howpublished={\url{https://huggingface.co/blog/lightonai/denseon-lateon}},
}

@article{zhang2024mgte,
  title={mGTE: Generalized Long-Context Text Representation and Reranking Models for Multilingual Text Retrieval},
  author={Zhang, Xin and Zhang, Yanzhao and Long, Dingkun and Xie, Wen and Dai, Ziqi and Tang, Jialong and Lin, Huan and Yang, Baosong and Xie, Pengjun and Huang, Fei and Zhang, Meishan and Li, Wenjie and Zhang, Min},
  journal={arXiv preprint arXiv:2407.19669},
  year={2024}
}