sci_index_train

A high-quality semi-synthetic training set for embedding models for searching scientific literature. It consists of 142,478 training and 7,516 dev queries over arXiv, with 631,611 query-paper positive pairs and mined hard negatives.

We create this dataset to model how users actually pose queries for literature search. We mostly target descriptive information needs ("recent advances in formal verification of stochastic dynamical systems") instead of titles or keywords. Many of these queries have several relevant papers, not one. These are the cases where general-purpose embedding models do worst.

We employ this dataset to fine-tune granite-embedding-small-english-r2 in our scientific_granite model. On the held-out benchmark sci_index_bench it improves ndcg@10 from 0.756 to 0.800, with +0.082 on exploratory queries, showing the value of this dataset for fine-tuning embedding models for literature search. The gains also generalize to other benchmarks with gains in ndgc@10 of +0.057 on LitSearch over the base model.

Dataset composition

We synthetically sample different types of input queries.

category	rows	row %	positive pairs	pair %	pos/row	neg/row	median words	of which jargon variants
nl_question	48,864	34.3%	48,864	7.7%	1.0	14.5	13	6,477
exploratory	39,103	27.4%	389,344	61.6%	10.0	90.5	12	7,056
problem	27,172	19.1%	27,172	4.3%	1.0	14.3	8	2,759
cit_rewrite	15,579	10.9%	154,471	24.5%	9.9	89.2	21	105
known_item	5,994	4.2%	5,994	0.9%	1.0	14.7	12	966
keyword	5,766	4.0%	5,766	0.9%	1.0	8.3	5	1,640
train total	142,478		631,611					19,003

dev: 7,516 queries in the same proportions (nl_question 2,606 · exploratory 2,074 · problem 1,442 · cit_rewrite 751 · known_item 330 · keyword 313).

These categories simulate the following:

category	what it simulates	example
exploratory	a survey-style topic need with several relevant papers	current limitations and scaling challenges in high-density associative memory architectures
cit_rewrite	a specific research question, phrased the way other papers describe the cited work	How do latent diffusion models utilize a gradual spreading of data distribution to perform short-term rainfall prediction?
nl_question	a natural-language question about one paper's finding	Can federated learning match centralized training on generalizing to unseen mammographic data?
problem	a terse, search-engine-style description of a problem or method	domain expert annotated dataset for medical summarization
keyword	keyword fragments	improved sugama collision
known_item	a half-remembered paper: partial or garbled title, often with an author	a paper by Antamanova, something about Anode current saturation of

• Jargon variants (13.3% of rows, generator starting with jargon_) are copies of another query in the same category with a technical term swapped for its acronym, its expansion or a synonym. They are paired with the original query and share its labels, which teaches invariance to how a concept is named. Either a curated acronym lexicon (jargon_lexicon) or an LLM (jargon_llm) made the substitution.
• cit_rewrite queries are LLM paraphrases, written as questions, of sentences in which other arXiv papers cite the target paper. The sentences were retrieved via the Semantic Scholar API. No citation sentence is included verbatim.

Schema

column	type	notes
query_id	string	content hash of the query
category	string	query type (table above)
generator	string	specific generator; jargon_* marks a jargon variant
model_tier	string	model (or algorithmic) that produced q
register	string	question · brief · keyword · statement · recollection
specificity	string	broad · specific
paper_id	string	arXiv ID of the paper the query was generated from
q	string	the query text
positives	list[string]	arXiv IDs of relevant papers, sorted by positive_scores (descending)
positive_scores	list[float]	cross-encoder relevance in [0, 1], aligned to positives
hard_negatives	list[string]	arXiv IDs of mined non-relevant papers
no_in_batch_negatives	bool	true for multi-positive rows. Mask their positives from each other
dev_hard	bool	dev only: the lowest-lexical-overlap tercile (2,929 rows)

Single- vs multi-positive rows. For the four single-positive categories, positives == [paper_id]. For exploratory and cit_rewrite, positives is the top 10 candidates by cross-encoder score, whether or not they include the source paper. The source paper is in the set for 40% (exploratory) and 67% (cit_rewrite) of rows, because another paper often answers the query better. Treat positive_scores as soft labels. The lowest-ranked positives can score close to 0, so a hard 0/1 reading over-trusts the tail.

Positives and negatives are arXiv identifiers, not text. A script that rebuilds the paper text from the public arXiv metadata snapshot ships with the dataset (Getting the paper text).

Usage

We only include the arXiv identifiers with this dataset since not all papers on arXiv come with sufficiently permissive licenses. To use the dataset, one thus has to re-recreate the corpus from the kaggle snapshot of arXiv metadata. Download this snapshot, then you can use the rebuild_corpus.py script to recreate the required corpus:

pip install pyarrow

python rebuild_corpus.py --snapshot arxiv-metadata-oai-snapshot.json \
    --ids . --verify doc_checksums.parquet --out corpus.parquet

⚠️ Important Notes

1. Row share is not signal share. exploratory and cit_rewrite rows carry ~10 graded positives and ~90 hard negatives each. The other four categories carry one positive and at most 15 negatives. exploratory is 27% of rows but 62% of positive pairs. If you flatten rows to (query, positive) pairs, the mix you train on is not the mix in the row counts.
2. Mask positives out of in-batch negatives. Rows with no_in_batch_negatives = true (all exploratory and cit_rewrite rows) have several genuine positives. If two of a query's positives land in the same batch, standard MNRL will push them apart.
3. Don't tune on dev and then report sci_index_bench. All 6,825 benchmark queries come from this dataset's dev split. Use the benchmark's selection split for checkpoint selection and report its test split. dev here suits in-training monitoring.

How the dataset was built

1. Corpus. All arXiv papers in the July 2026 Kaggle metadata snapshot with a title and an abstract of at least 200 characters, deduplicated by normalised title: 3,040,442 papers.
2. Source papers. A stratified sample of source papers. Papers from the evaluation benchmarks below were excluded from the sample.
3. Query generation. Per category, LLM or algorithmic (tables above). The LLM prompts (shared by every LLM category, including the Gemini one) ask for a search need rather than a summary of the paper. They forbid naming the paper or its authors, never reuse the title as a phrase, and copy at most 4 consecutive words from the title or abstract.
4. Quality control. Schema validation, exact-normalised deduplication across all categories, fastText language ID, and a check that no keyword / known_item query is simply the paper's title. All categories except exploratory also pass a round-trip filter, which keeps a query only if a frozen granite-embedding-311m-multilingual-r2 ranks its source paper in the top 100 of the full corpus.
5. Hard-negative mining. Candidates are the union of BM25 and dense retrieval (top 100 each) over the full corpus.
   • Single-positive rows: NV-Retriever TopK-PercPos drops any candidate scoring ≥ 95% of the positive. A BAAI/bge-reranker-v2-m3 cross-encoder then drops any negative scoring within 0.05 of the positive. At most 15 negatives are kept.
   • Multi-positive rows: the cross-encoder scores every candidate. The top 10 become graded positives, candidates below the median score become negatives, and the ambiguous middle band is discarded.
6. Packaging. Rows left with no negatives are dropped. For categories over quota, the 5% tails of the positive–negative cross-encoder margin are trimmed before random downsampling. exploratory has no single positive to take a margin against, so it is only downsampled. The train/dev split is by source paper (5% dev).
7. Evaluation-set decontamination. Every paper in the LitSearch and DORIS-MAE corpora that could be matched to arXiv (by ID for DORIS-MAE, by normalised title for LitSearch) is removed from every role: source papers, positives and hard negatives. That is 20,898 papers in total, and it removed 3,005 positives and 46,218 negatives from training rows. Six rows were left with no negatives and dropped. The released data references none of the 20,898 papers (verified).

Limitations

• Synthetic queries. All queries are LLM-generated or algorithmically constructed, not sampled from search logs. They are designed to resemble literature-search queries and are not evidence of what people actually ask. The LitSearch and DORIS-MAE results are evidence that training carries over to human-written queries but we expect possible gains from actual human logs.
• Uneven query-shape coverage. keyword + known_item are 8.2% of rows, and title-as-query is absent. Models trained on this data regress on title lookup and keyword-style benchmarks.
• Uneven negative depth. Multi-positive rows average ~90 negatives. The rest are capped at 15, and keyword averages 8.3 (37.5% of keyword rows have fewer than 5).
• The train/dev split is by source paper only. A paper can still appear as a positive on both sides: 32.7% of papers that are dev positives are also train positives. No query text or source paper is shared. sci_index_bench removes this overlap from its relevance labels.
• Labels come from models. Multi-positive sets and negative filtering rely on a cross-encoder. There are no human relevance judgements.
• Deduplication is exact-normalised only. There is no near-duplicate pass. Jargon variants are near-duplicates by design.
• English and arXiv only. Physics, mathematics and computer science are well covered. Biomedicine, chemistry and the social sciences are thin.
• No test split. Report on sci_index_bench test or an external benchmark.

Licence

MIT. The dataset contains generated query text and arXiv identifiers. No paper titles, abstracts or full texts are redistributed. The arXiv metadata compilation it refers to is CC0.