NSR Reranker

Cross-encoder reranking over nuclear-physics literature (277,068 Nuclear Science References).

✅ More than doubles top-1 on expert queries: re-scoring stock bge-m3's top-50, R@1 0.080 → 0.180 (2.25×), nDCG@10 +76%.
✅ Recovers 97% of what the pool contains: those 50 candidates hold the right paper 24.9% of the time; this model surfaces it into the top ten 24.2% of the time.
✅ Better on every segment — all 8 segments, both query families, on R@10 and nDCG@10 (pre-registered gate, passed).
✅ Drop-in: standard sentence-transformers CrossEncoder, 149M parameters, Apache-2.0.

A 149M cross-encoder fine-tuned from Alibaba-NLP/gte-reranker-modernbert-base on expert-written query→paper groups from the NSR corpus. Takes (query, document), returns a relevance logit. Second stage over any dense or hybrid top-K.

Second stage of a two-stage stack — pair it with the NSR Encoder for best results.

Details

Property	nsr-reranker
Type	Cross-encoder (pointwise reranker)
Total parameters	149M
Backbone	Alibaba-NLP/gte-reranker-modernbert-base (ModernBERT)
Output	Single relevance logit per `(query, document)` pair
Context	8,192 tokens native; trained and served at 256
Training signal	39.6k expert query→paper groups, 7 hard negatives each
Built for	Second-stage reranking of nuclear-physics search results
Pair with	NSR Encoder (first stage)
License	Apache-2.0

Performance

Expert keyword queries (KW, n = 4,998) — an NSR indexer's structured keyword abstract as the query. Benchmark NSR Eval, frozen before any training, split by paper; candidate pools retrieved against all 277,068 documents.

What the second stage buys — stock bge-m3's top-50, before and after this model:

Metric	Pool alone	+ this reranker	Δ
R@1	0.080	0.180	+125%
R@10	0.171	0.241	+41%
nDCG@10	0.121	0.213	+76%

By segment — nDCG@10, the same top-50 before and after:

Segment	n	Pool alone	+ this reranker	Δ
title-only documents	4,141	0.085	0.167	+96%
has-abstract documents	857	0.296	0.436	+47%
pre-1970	602	0.045	0.107	+137%
1970–1999	2,686	0.105	0.197	+89%
2000+	1,710	0.174	0.274	+58%
journal articles	4,250	0.129	0.221	+71%
other reference types	748	0.075	0.165	+119%

EXFOR queries (EX, n = 4,997) — on an already-saturated pool it still moves R@1 from 0.871 to 0.927 (nDCG@10 0.878 → 0.912) and is the strongest arm outright.

Ceiling. A reranker cannot retrieve what its candidate list does not contain. The pool measured here contains the answer 24.9% of the time; in production this model sits over RRF(FTS + NSR Encoder), whose top-50 contains it 65.6% of the time.

Training


Objective	grouped cross-entropy — 1 positive vs 7 hard negatives per group
Groups	39,600 — NSR keyword-abstract and EXFOR queries; positive = the annotated paper
Negatives	mined from the retrieval space itself (base-model kNN + a lexical arm), not sampled at random
Epochs / max length / lr	2 / 256 / 2e-5
Hardware / wall-clock	1× NVIDIA L40S (`g6e.xlarge`) · 42.6 min
Run	`20260817-032126-train` · 2026-08-17

Split by paper; every benchmark paper is excluded as a query source and positive.

How to run

from sentence_transformers import CrossEncoder

ce = CrossEncoder("NYSgpt/nsr-reranker", max_length=256)
ce.predict([("92Zr(n,γ) cross section",
             "Neutron capture cross sections of 92Zr and their astrophysical implications ...")])

Retrieve a top-50 with the NSR Encoder, re-score with this model, serve the reordered list.

📬 Contact

Questions, results, or a use case to share? Open a discussion in the Community tab.

Citation

@misc{nsrreranker2026,
  title  = {NSR Reranker: a cross-encoder for the Nuclear Science References corpus},
  author = {NYSgpt},
  year   = {2026},
  url    = {https://huggingface.co/NYSgpt/nsr-reranker}
}