daniel-gomm/sci_index_bench

SciIndexBench v1

A synthetically generated retrieval benchmark generated with the same generation pipeline for queries as the science-index training dataset. The benchmark aims for more natural and varied search queries for identifying relevant papers against paper abstracts from arxiv.

We release the benchmark alongside the training dataset to use for benchmarking text embedding models on paper retrieval. It is completely decoupled from the science-index training dataset, sharing no source and gold papers.

Splits

split	queries	purpose
selection	2,043	checkpoint selection / early stopping / tuning
test	4,782	reporting

The selection and the test split are completely independent, not sharing any source or gold papers. test is used to report final results.

Contents

queries	6,825
qrels	22,672 (3.322 gold papers per query)
corpus	316,646 arXiv papers (300,000 uniform random + 18,441 gold)
hard slice	queries flagged dev_hard sit in the lowest lexical-overlap tercile

The dataset is made up of different categories of queries:

query type	n	share
nl_question	2,301	33.7%
exploratory	2,045	30.0%
problem	1,191	17.5%
cit_rewrite	751	11.0%
known_item	296	4.3%
keyword	241	3.5%

Baselines

On the selection split, binary relevance, exact cosine search over the full corpus.

model	ndcg@10	recall@10	recall@100	mrr@10
granite-embedding-small-english-r2 (out-of-the-box, 47M, 384d)	0.7535	0.8074	0.9491	0.8315

Usage

We do not redistribute contents of arxiv papers as not all papers have sufficiently permissive licenses. Instead, the document text has to be rebuilt by downloading a copy of the Kaggle arXiv snapshot and running the python script rebuild_corpus.py provided alongside the dataset. To run the script, make sure that pyarrow is installed (e.g., by running pip install pyarrow).

Prerequisite for rebuilding the corpus is

python rebuild_corpus.py --snapshot arxiv-metadata-oai-snapshot.json \\
        --ids path/to/dataset_dir --ids path/to/corpus_ids.json --out corpus.parquet

import json, polars as pl

queries = pl.read_parquet("queries.parquet").filter(pl.col("split") == "test")
qrels   = pl.read_parquet("qrels.parquet")
corpus_ids = set(json.load(open("corpus_ids.json")))

docs = pl.read_parquet("corpus.parquet").filter(pl.col("id").is_in(list(corpus_ids)))
# document template used throughout: "Title:\n{title}\nAbstract:\n{abstract}"

Limitations

• Queries are synthetic — LLM-generated or algorithmically constructed from a paper's own metadata, not sampled from user logs. They are built to resemble literature-search queries; they are not evidence of what researchers ask.
• Easier than LitSearch. An out-of-the-box granite-embedding-small-english-r2 scores ndcg@10 ≈ 0.75 here versus ≈ 0.49 on LitSearch, because a query derived from a paper's own abstract is more findable than a real human query. Use the dev_hard slice for a harder read.
• English and arXiv only, with arXiv's field distribution: strong in CS physics and maths; no clinical medicine, chemistry, life sciences or social science.
• Shared corpus across splits by design, so selection and test numbers are directly comparable.