lightonai/embeddings-fine-tuning-filtered-en

Overview

This dataset is composed of high quality data sources with mined hard negatives annotated with bi-encoder and cross-encoder scores. It can be used to train a strong retrieval model by itself but is better used after a large-scale contrastive pre-training, for example using our multilingual and English datasets. The negatives were mined following the NV-Retriever setup: the closest documents to each query are mined as negatives, and false negatives are filtered out if their bi-encoder similarity is higher than a percentage of the query-positive similarity score. This dataset is a ready-to-train filtered version of embeddings-fine-tuning, extended with the English MIRACL and MLDR data filtered from embeddings-fine-tuning-multilingual-unfiltered: we keep the 10 hardest negatives per sample after NV-Retriever filtering with a threshold of 0.95, and remove the samples with less than 10 valid negatives as they may contain weakly annotated pairs. The mined datasets are FiQa, NaturalQuestion, HotpotQA, MSMARCO, FEVER, SquadV2, TriviaQA, MIRACL and MLDR, each sample containing the query, the positive and 10 mined hard negatives. The model used for mining is gte-modernbert-base, except for MIRACL and MLDR where snowflake-arctic-embed-l-v2.0 was used, and all the samples were annotated with the cross-encoder mxbai-rerank-large-v2, enabling knowledge distillation training on top of contrastive learning.

For more information, please read our multilingual models blog post, our English models blog post and our paper.

How to use

The negatives are already mined and filtered, so using the data as contrastive data in either sentence-transformers or PyLate only requires joining the three subsets into the (query, positive, negative_0, negative_1, ..., negative_n) format. The cross-encoder rerank_scores of the kept documents are carried along in the same order as the columns, so they can be used as teacher scores by a knowledge distillation loss (a KL-divergence between the student and teacher relevance distributions, for instance) on top of the contrastive loss:

Python code to cast to contrastive format

import datasets


class KDToContrastive:
    """Maps the scores table of a split to the contrastive knowledge distillation format.

    Parameters
    ----------
    queries
        Queries subset of the split.
    documents
        Documents subset of the split.
    num_negatives
        Number of hard negatives to keep per query, out of the 10 stored ones.
    """

    def __init__(
        self,
        queries: datasets.Dataset,
        documents: datasets.Dataset,
        num_negatives: int = 10,
    ) -> None:
        self.queries = dict(zip(queries["query_id"], queries["query"]))
        self.documents = dict(zip(documents["document_id"], documents["document"]))
        self.num_negatives = num_negatives

    def map_to_query_positive_negatives(self, example) -> dict:
        # document_ids, scores and rerank_scores are all ordered [positive, negative_0, ..., negative_9]
        document_ids = example["document_ids"][: self.num_negatives + 1]
        return {
            "query": self.queries[example["query_id"]],
            "positive": self.documents[document_ids[0]],
            "teacher_scores": example["rerank_scores"][: self.num_negatives + 1],
            **{
                f"negative_{negative}": self.documents[document_id]
                for negative, document_id in enumerate(document_ids[1:])
            },
        }


def load_train_datasets(num_negatives: int = 10) -> datasets.DatasetDict:
    """Load every split as a (query, positive, negatives, teacher_scores) dataset."""
    repo = "lightonai/embeddings-fine-tuning-filtered-en"
    splits = [
        "fiqa_en", "hotpotqa_en", "nq_en", "msmarco_en", "fever_en", "squadv2_en", "trivia_en",
        "miracl_en", "mldr_en",
    ]

    train_dataset = datasets.DatasetDict()
    for split in splits:
        # data_files restricts the download to the split being processed, hence skipping the checks on the other splits
        load = lambda config: datasets.load_dataset(
            repo,
            name=config,
            data_files=f"{config}/{split}-*",
            split="train",
            verification_mode="no_checks",
        )
        scores = load("scores")
        processor = KDToContrastive(
            queries=load("queries"), documents=load("documents"), num_negatives=num_negatives
        )
        train_dataset[split] = scores.map(
            processor.map_to_query_positive_negatives,
            remove_columns=scores.column_names,
            desc=f"Creating the contrastive dataset ({split})",
        )
    return train_dataset


train_dataset = load_train_datasets()
print(train_dataset)

Dataset structure

The dataset is composed of 9 high quality datasets, defined by the splits parameters. Each split contains 3 subsets, one containing the queries, one containing the documents and one joining tables also containing the corresponding pairwise query-documents scores.

Documents

Column	Type	Description
document_id	int64	Unique identifier of the document within the split.
document	string	Raw text of the document/passage.

Split	Rows
fiqa_en	31.9k
hotpotqa_en	830k
nq_en	896k
msmarco_en	3.63M
fever_en	210k
squadv2_en	19k
trivia_en	2.76M
miracl_en	59.1k
mldr_en	53.8k
Total	8.50M

Queries

Column	Type	Description
query_id	int64	Unique identifier of the query within the split.
query	string	Raw text of the query.

Split	Rows
fiqa_en	5.4k
hotpotqa_en	84k
nq_en	114k
msmarco_en	492k
fever_en	110k
squadv2_en	129k
trivia_en	56.7k
miracl_en	2.6k
mldr_en	9.1k
Total	1.00M

Scores

Column	Type	Description
query_id	int64	Identifier joining back to the corresponding row in queries.
document_ids	list[int64]	List of document IDs (joining back to documents). The first element is the positive document, followed by the 10 hardest negatives kept after NV-Retriever filtering.
scores	list[float]	Bi-encoder relevance scores for each document w.r.t the query, in the same order as document_ids. Can be used for knowledge distillation.
rerank_scores	list[float]	Cross-encoder scores from mxbai-rerank-large-v2 for each document w.r.t the query, in the same order as document_ids. Can be used for knowledge distillation.

Split	Rows
fiqa_en	13k
hotpotqa_en	143k
nq_en	114k
msmarco_en	519k
fever_en	127k
squadv2_en	129k
trivia_en	511k
miracl_en	6.2k
mldr_en	9.1k
Total	1.57M

Token length distributions

Token counts are computed with the mmBERT-base tokenizer. For readability, each histogram is truncated after the last bin containing at least 5 samples; the statistics reported in the boxes (including the maximum) are computed on the full data.

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},
}```