NYSgpt/nsr-eval

NSR Eval

A frozen retrieval benchmark over nuclear-physics literature (277,068 Nuclear Science References).

NSR Collection | NSR Encoder | NSR Reranker

The benchmark behind every number on the NSR Encoder and NSR Reranker cards: the query keys, the gold labels, and the results of every arm we scored — so anyone can check a published number or score their own model on the same queries.

• ✅ Frozen before any training, split by paper. 25% of papers held out by a declared hash rule; no benchmark paper is a query source or positive in training.
• ✅ Two expert query families. Indexer-written keyword abstracts (KW) and EXFOR experiment entries (EX) — real professional-search queries, not synthetic questions.
• ✅ Every baseline measured here, not quoted. Same corpus, same queries, same metric code; reported per segment (era, document richness, reference type).
• ✅ Keys-only. NSR key numbers and EXFOR entry numbers, not NNDC/IAEA text — every row resolves at its public source.

---

Contents

file	rows	what
queries.jsonl	9,995	{qid, family, source_scheme, source_id, source_url, era, richness, reftype} — the key of the record the query is built from, plus segment tags
qrels.tsv	11,048	qid, key_number, url — the gold NSR paper(s); EX queries may have several
manifest.json	1	freeze time, holdout rule, per-cell counts, leak audit
results/*.json	6 arms	metrics per family and per segment
results/bge-m3+nsr-reranker.perquery.json	9,995	the stock top-50 and the reranked top-50 as key numbers, per query

Query families

family	queries	query is	gold is
KW	4,998	the NSR keyword abstract of paper source_id — the indexer's structured REACTION / MEASURED / DEDUCED phrases, flattened to one line	that same paper
EX	4,997	EXFOR entry source_id — its title and reaction codes (e.g. 64-GD-0(N,TOT),,SIG)	the NSR paper(s) the entry is linked to

Rebuilding the query text. KW: fetch the NSR record at source_url and join its keyword-abstract phrases with ; . EX: fetch the EXFOR entry at source_url and join its title with its reaction strings, ; -separated. Both sources are public databases.

---

Performance

Expert keyword queries (KW, n = 4,998), retrieved against all 277,068 documents.

Rank	Arm	R@1	R@10	nDCG@10	file
1	RRF(FTS + NSR Encoder)	0.344	0.542	0.437	rrf_fts+nsr-encoder
2	NSR Encoder	0.252	0.487	0.363	nsr-encoder
3	RRF(FTS + stock bge-m3)	0.212	0.282	0.244	rrf_fts+bge-m3
4	stock bge-m3 + NSR Reranker	0.180	0.241	0.213	bge-m3+nsr-reranker
5	Postgres FTS	0.161	0.165	0.163	fts
6	BAAI/bge-m3 (stock)	0.080	0.171	0.121	bge-m3

EXFOR queries (EX, n = 4,997) — near-saturated for every dense arm: stock bge-m3

• NSR Reranker 0.927 R@1, NSR Encoder 0.911, stock bge-m3 0.871, FTS 0.017.

Every results/*.json carries the same metrics per segment: era (pre1970, 1970-1999, 2000+), richness (has-abstract, title-only), reftype (journal, other). Read per segment — the two families flip which arm wins, and a blended row alone hides it.

---

Details

Property
Corpus	277,068 NSR references with embeddable text (title + publisher abstract when present, title only otherwise); NNDC NSR snapshot 2026-06-30
Unit of record	one NSR reference, identified by its key_number
Split	25% of papers held out by `md5("holdout:"
Selection	stratified across era × richness cells; 4,998 KW + 4,997 EX
Metrics	R@1, R@5, R@10, R@50, MRR, nDCG@10 — per family and per segment
Twins	181 KW positives have an exact-text twin in the corpus; a twin hit is scored as correct
Frozen	2026-08-17T01:42Z

Leak audit at freeze time: the embedded document is title + publisher abstract, the keyword abstract lives in a separate field — 0 of 9,995 queries appear inside their own document.

---

How to use

import json, csv

queries = [json.loads(l) for l in open("queries.jsonl")]
gold = {}
for r in csv.DictReader(open("qrels.tsv"), delimiter="\t"):
    gold.setdefault(r["qid"], set()).add(r["key_number"])

# score your model: for each query, the rank of the first gold key_number in your top-50
hits = json.load(open("results/bge-m3+nsr-reranker.perquery.json"))  # {qid: {raw_top50, reranked}}

---

Licensing

NSR keyword abstracts are written by NNDC indexers and EXFOR entries are compiled by the IAEA/NNDC network; both are third-party text and neither is redistributed here. This release ships keys — NSR key numbers, EXFOR entry numbers, segment tags — and our own measurements, released CC-BY-4.0. The text behind each key stays at its source, under its source's terms.

---

📬 Contact

Questions, results, or a model to add to the table? Open a discussion in the Community tab.

Citation

@misc{nsreval2026,
  title  = {NSR Eval: a frozen retrieval benchmark over the Nuclear Science References corpus},
  author = {NYSgpt},
  year   = {2026},
  url    = {https://huggingface.co/datasets/NYSgpt/nsr-eval}
}