hotchpotch/bekko-embedding-v1-hard-negatives

hotchpotch/bekko-embedding-hard-negatives

hotchpotch/bekko-embedding-hard-negatives is a private collection of hard-negative training subsets for information retrieval and embedding-model training. Each row contains one query, one positive text, and 15 mined hard negatives in a flat schema:

query, pos, neg_1, neg_2, ..., neg_15

The dataset is intended for training dense retrievers, sparse/dense hybrid retrievers, embedding models, and rerankers that benefit from explicit challenging negatives. The collection currently covers English web text, biomedical abstracts, arXiv papers, code-search pairs, English Wikipedia paragraph retrieval, multilingual Wikipedia long-document retrieval, and several Sentence Transformers query-pair datasets.

The dataset card documents the durable upstream datasets and construction methods for the current hard-negative subsets. Some intermediate hard-negative datasets used during assembly may be removed; use this dataset and the upstream source references below as the stable references.

Loading

from datasets import load_dataset

ds = load_dataset(
    "hotchpotch/bekko-embedding-hard-negatives",
    "natural_questions",
    split="train",
)
print(ds[0].keys())

For multilingual Wikipedia long-document hard negatives:

from datasets import load_dataset

ds = load_dataset(
    "hotchpotch/bekko-embedding-hard-negatives",
    "wikipedia_hard_negatives_long_docs_ja",
    split="train",
)

Dataset Structure

All subsets expose the same columns.

Column	Type	Description
query	string	The search query, premise, caption, news title, or retrieval query depending on the source subset.
pos	string	The positive retrieval target for the query.
neg_1 ... neg_15	string	Mined hard negative texts from the same source family or retrieval corpus.

The exported rows do not include document IDs, scores, rank positions, query-generation metadata, source labels, or mining diagnostics. If you need provenance-aware training or evaluation, keep a separate local mapping to the original source datasets.

Subsets

Subset	Rows	Language	Positive text type	Stable upstream reference	Query source / generation	Mining method	Recommended use
fineweb	20,000	English	Web document text	hotchpotch/fineweb-ir-simulated-search-queries, derived from HuggingFaceFW/fineweb-edu	Simulated queries generated by the upstream dataset	Hybrid BM25 + dense search	General web retrieval training
pubmed	20,000	English	PubMed title + abstract	hotchpotch/pubmed-abstract-ir-simulated-search-queries, derived from slinusc/PubMedAbstractsSubset	Simulated specialist search queries generated by the upstream dataset	Hybrid BM25 + dense search	Biomedical retrieval training
arxiv	20,000	English	arXiv title + abstract	hotchpotch/arxiv-ir-simulated-search-queries, derived from open-index/open-arxiv	Simulated specialist search queries generated by the upstream dataset	Hybrid BM25 + dense search	Scientific retrieval training
codesearch	20,000	English / code	Code snippet or code-related document	nomic-ai/nomic-embed-unsupervised-data, codesearch subset	Weakly supervised code-search pairs released with Nomic Embed unsupervised data	Hybrid BM25 + dense search	Code and technical retrieval training
natural_questions	20,000	English	Wikipedia-style answer passage	sentence-transformers/natural-questions, pair subset	Human/web question-answer pairs from the upstream dataset	Hybrid BM25 + dense search	Recommended QA retrieval subset
agnews	20,000	English	News text	sentence-transformers/agnews, pair subset	News title/text pairs from the upstream dataset	Hybrid BM25 + dense search	Recommended news retrieval subset
trivia_qa	20,000	English	TriviaQA evidence text	sentence-transformers/trivia-qa, pair subset	Trivia question/evidence pairs from the upstream dataset	Hybrid BM25 + dense search	Recommended QA retrieval subset
gooaq	20,000	English	Short answer or answer passage	sentence-transformers/gooaq, pair subset	Google-style question-answer pairs from the upstream dataset	Hybrid BM25 + dense search, regenerated with top-k 500	Recommended QA retrieval subset
all_nli	20,000	English	NLI hypothesis or sentence	sentence-transformers/all-nli, pair subset	Natural-language inference sentence pairs from the upstream dataset	Hybrid BM25 + dense search, regenerated with top-k 500	Use cautiously; negatives can be semantic near-positives
coco_captions	20,000	English	Image caption text	sentence-transformers/coco-captions, pair subset	Caption pairs from the upstream dataset	Hybrid BM25 + dense search, regenerated with top-k 500	Use cautiously or with low weight; caption negatives are often very close
wikipedia_hard_negatives_english	500,000	English	Wikipedia paragraph	hotchpotch/wikipedia-english-ir-simulated-search-queries	Simulated English search queries over Wikipedia paragraph positives	Hybrid BM25 + dense search	Large-scale English Wikipedia retrieval training
wikipedia_hard_negatives_long_docs_ar	4,218	Arabic	Wikipedia long document	hotchpotch/wikipedia-multilingual-synthetic-ir-query and Wikipedia-derived source text	Long-document query generation with deepseek-v4-flash	Qdrant sparse BM25 only	Multilingual long-document retrieval
wikipedia_hard_negatives_long_docs_de	4,825	German	Wikipedia long document	Same upstream as the Arabic long-document subset	Same generation method	Qdrant sparse BM25 only	Multilingual long-document retrieval
wikipedia_hard_negatives_long_docs_en	16,212	English	Wikipedia long document	Same upstream as the Arabic long-document subset	Same generation method	Qdrant sparse BM25 only	Multilingual long-document retrieval
wikipedia_hard_negatives_long_docs_es	4,932	Spanish	Wikipedia long document	Same upstream as the Arabic long-document subset	Same generation method	Qdrant sparse BM25 only	Multilingual long-document retrieval
wikipedia_hard_negatives_long_docs_fr	4,958	French	Wikipedia long document	Same upstream as the Arabic long-document subset	Same generation method	Qdrant sparse BM25 only	Multilingual long-document retrieval
wikipedia_hard_negatives_long_docs_it	4,846	Italian	Wikipedia long document	Same upstream as the Arabic long-document subset	Same generation method	Qdrant sparse BM25 only	Multilingual long-document retrieval
wikipedia_hard_negatives_long_docs_ja	3,676	Japanese	Wikipedia long document	Same upstream as the Arabic long-document subset	Same generation method	Qdrant sparse BM25 only	Multilingual long-document retrieval
wikipedia_hard_negatives_long_docs_ko	3,911	Korean	Wikipedia long document	Same upstream as the Arabic long-document subset	Same generation method	Qdrant sparse BM25 only	Multilingual long-document retrieval
wikipedia_hard_negatives_long_docs_pt	4,793	Portuguese	Wikipedia long document	Same upstream as the Arabic long-document subset	Same generation method	Qdrant sparse BM25 only	Multilingual long-document retrieval
wikipedia_hard_negatives_long_docs_ru	3,615	Russian	Wikipedia long document	Same upstream as the Arabic long-document subset	Same generation method	Qdrant sparse BM25 only	Multilingual long-document retrieval
wikipedia_hard_negatives_long_docs_zh	3,601	Chinese	Wikipedia long document	Same upstream as the Arabic long-document subset	Same generation method	Qdrant sparse BM25 only	Multilingual long-document retrieval

Total rows: 759,587.

ccnews is intentionally not included in this bundle because manual review found it less suitable than the retained hard-negative subsets.

Construction

Hybrid hard negatives

The fineweb, pubmed, arxiv, codesearch, natural_questions, agnews, trivia_qa, gooaq, all_nli, coco_captions, and wikipedia_hard_negatives_english subsets were mined with hybrid retrieval over Qdrant. The index combined sparse BM25 retrieval with dense vectors from an internal embedding model.

Setting	Value
Dense model	hotchpotch/bekko-embedding-v1-a8m-prelease (internal)
Dense vector truncation	256 dimensions
Dense vector dtype	bfloat16
BM25 tokenization	English regex tokenization
Post-processing	Positive-document exclusion, duplicate/near-duplicate exclusion, suspicious-negative filtering
Short-query overlap filter	Enabled for queries with at most 4 tokens; candidates with coverage >= 0.8 are filtered

Each exported row contains 15 negatives. Rank ranges below are inclusive integer ranges.

Subsets	Candidates retrieved per query	Negative rank buckets
fineweb, pubmed, arxiv, codesearch, natural_questions, agnews, trivia_qa	Top 100	Ranks 20-49: 7 negatives; ranks 50-99: 8 negatives
gooaq, all_nli, coco_captions	Top 500	Ranks 100-199: 7 negatives; ranks 200-499: 8 negatives
wikipedia_hard_negatives_english	Top 1000	Ranks 20-199: 7 negatives; ranks 200-999: 8 negatives

The deeper top-k policy for gooaq, all_nli, and coco_captions was chosen after manual review found shallow negatives too easy, too repetitive, or too close to the positive.

Wikipedia multilingual long-document hard negatives

The multilingual long-document subsets are derived from Wikipedia long documents. They follow the same broad idea as Shitao/MLDR: a paragraph or evidence span inside a long document is used to generate a retrieval-style query, and the whole long document is the positive target. MLDR's dataset card states that it uses GPT-3.5-generated questions; these subsets instead use DeepSeek-generated queries over Wikipedia-derived source text.

The long-document hard negatives were mined per language with Qdrant sparse BM25 only. The BM25 index used body-only document text with the title prefix removed before indexing and in the final exported text.

Setting	Value
Retrieval method	Sparse BM25 only
Candidates retrieved per query	Top 500
Positive handling	Positive document excluded before sampling
Negative rank buckets	Ranks 100-199: 7 negatives; ranks 200-499: 8 negatives
Tokenization	Japanese, Korean, and Chinese use language-specific tokenization; other languages use regex tokenization

Rows that could not provide all 15 filtered negatives were skipped.

Mining Diagnostics

Subset	Output rows	Candidate rows seen	Suspicious negatives skipped	Insufficient negatives skipped
fineweb	20,000	25,928	5,928	0
pubmed	20,000	20,063	63	0
arxiv	20,000	20,121	121	0
codesearch	20,000	22,052	2,050	2
natural_questions	20,000	26,672	6,672	0
agnews	20,000	23,315	3,315	0
trivia_qa	20,000	26,644	6,644	0
gooaq	20,000	28,070	8,070	0
all_nli	20,000	20,424	424	0
coco_captions	20,000	25,865	5,865	0

Query Generation Sources

The non-Wikipedia 20k source datasets are based on upstream query-pair datasets whose dataset cards should be treated as the primary source of provenance:

• FineWeb: simulated English search queries for FineWeb-Edu-derived web documents.
• PubMed: simulated specialist biomedical search queries for PubMed title-and-abstract records.
• arXiv: simulated specialist scientific search queries for arXiv title-and-abstract records.
• Nomic codesearch: weakly supervised contrastive text-pair data from nomic-ai/nomic-embed-unsupervised-data, codesearch subset.
• Sentence Transformers datasets: query-pair or sentence-pair data from the corresponding sentence-transformers/* upstream datasets.
• English Wikipedia paragraphs: simulated English search queries over Wikipedia paragraph positives.

The multilingual Wikipedia long-document subsets use this query-generation prompt pattern:

You generate one query for training an information retrieval model for long-document retrieval.

The input is a paragraph selected from a long document. Write one useful search query whose answer is fully supported by that paragraph and whose wording helps retrieve the source document.

Requirements:
- Prefer the central entity, event, mechanism, relation, or claim of the paragraph over a minor side fact.
- The query must be specific enough to retrieve the source document, not a broad question that could match many unrelated documents.
- The query must include at least one paragraph-specific anchor, such as a named entity, work title, place, date, number, technical term, law, event, or organization.
- The query must ask about exactly one topic.
- The query must contain only one question.
- The query must be answerable using only the provided paragraph.
- Do not write a yes/no question; ask for a concrete entity, date, number, term, relationship, function, role, value, action, or explicitly stated explanation.
- Before writing the final query, choose an answer target from the paragraph. If that target looks malformed, mistranslated, noisy, semantically inconsistent, peripheral, or dependent on missing notation, reject it and choose a clearer central target.
- Preserve the exact spelling of names, titles, organizations, technical terms, and quoted terms from the paragraph.
- Prefer asking about a stated function, role, value, action, or relation over asking a broad definition such as "What is X?".
- When asking for a number, date, percentage, or count, make the measured subject explicit in the query.
- Do not ask for causes, effects, implications, background, proof, or interpretation unless the paragraph explicitly states them.
- If the paragraph contains broken formulas, missing symbols, placeholders, or unclear notation, do not ask about the missing notation; ask about a clearly stated named entity, date, number, or term instead.
- Do not make an article, source, paragraph, passage, or text the subject of the question.
- Do not mention "the text", "the paragraph", "the passage", or similar source references.
- Do not use pronouns such as "it", "this", "that", "these", or "those".
- Make the subject of the query explicit enough that the query remains meaningful outside the paragraph.
- Keep the query concise, grammatical, and natural.
- End the query with a question mark or the standard question punctuation for the output language.
- Output only the query. Do not include explanations, labels, quotes, or additional content.

Write the query in "{language name}".

Intended Use

This dataset is suitable for:

• training dense retrieval and embedding models with explicit hard negatives;
• training rerankers or cross-encoders that consume query, pos, and multiple negatives;
• stress-testing retrieval models on lexically or semantically close negatives;
• mixing domain-specific hard-negative data into a broader retrieval training recipe;
• multilingual long-document retrieval experiments using Wikipedia-derived positives and BM25-only hard negatives.

Quality Notes

Manual review of the recently added Sentence Transformers hard-negative subsets suggests the following practical grouping:

• Strong general-purpose additions: natural_questions, agnews, trivia_qa, and gooaq.
• More specialized or lower-weight additions: all_nli and coco_captions.

all_nli contains short entailment-style pairs. Some mined negatives are semantically close to the positive, which can be useful for fine-grained semantic training but risky if treated as ordinary retrieval negatives. coco_captions contains caption-like texts where many examples describe visually similar scenes; this makes the negatives hard, but it can also create near-positive ambiguity. Use these subsets with lower sampling weight if your training objective assumes clearly non-relevant negatives.

License and Attribution

This aggregate follows the licenses and terms of the upstream datasets used to build each subset. The Hugging Face dataset metadata is preserved in this card; users should also check each upstream dataset card before redistribution or commercial use. Wikipedia-derived subsets should be treated as Wikipedia-derived content and are subject to Wikipedia's attribution and share-alike requirements, including CC BY-SA-style obligations. PubMed, arXiv, FineWeb, Nomic, and Sentence Transformers source subsets may have different license or usage constraints depending on their upstream data.