license: cc-by-4.0
task_categories:
- text-ranking
- text-retrieval
language:
- en
tags:
- benchmark
- reranker
- hard-negatives
- peer-review
- scientific-documents
- information-retrieval
pretty_name: 44b Eval
size_categories:
- 1K<n<10K
44b Eval
A frozen 9-way reranking benchmark over machine-learning papers, queried by their peer reviews.
The benchmark behind every number on the 44b reranker cards: which review is the query, which paper is the gold, which eight papers are the hard negatives, and the scored result of every arm — so anyone can check a published number or score their own reranker on the same 3,157 queries.
- ✅ Paper-disjoint from training. None of the 1,534 test papers appears in training as a positive; 23 (1.5%) appear as some other query's negative, which biases against the fine-tuned models rather than for them.
- ✅ Real expert queries. Each query is a peer reviewer's own summary of a paper, written by a domain expert on OpenReview.
- ✅ Hard by construction. 8 negatives per query mined from embedding-space nearest neighbours and lexical near-matches, screened for duplicates. Chance P@1 = 0.111.
- ✅ Keys-only. OpenReview note IDs and arXiv / OpenAlex / Semantic Scholar paper IDs, each with its public URL — no review text, no abstracts.
Contents
| file | rows | what |
|---|---|---|
test.jsonl |
3,157 | {note_id, note_url, kind, positive{scheme,id,url}, negatives[8]{scheme,id,url}} — one line per query |
results.jsonl |
7 | the scorer's output per arm: P@1, P@2, P@3, MRR, mean rank, ms/query, truncation settings |
manifest.json |
1 | split rule and seed, counts, disjointness proof, negative-mining recipe, ID-scheme census |
Rebuilding the text. The query is the summary field of an official_review (or the
metareview of a meta_review) at note_url. Each candidate is the title + abstract of
the paper at its url: arxiv IDs resolve at arxiv.org, openalex IDs at openalex.org,
s2 IDs at semanticscholar.org. 23 negatives (0.1%) are curated web documents with no
third-party identifier and are marked resolvable: false.
Performance
3,157 queries over 1,534 papers, 9 candidates each (1 positive + 8 hard negatives). Model arms on one A10G; retrieval arms are first-stage rankings of the same 9-candidate sets.
| Rank | Arm | P@1 | P@3 | MRR | mean rank | ms/query |
|---|---|---|---|---|---|---|
| 1 | 44b-reranker (149M, fine-tuned) | 0.929 | 0.982 | 0.956 | 1.148 | 41.1 |
| 2 | 44b-reranker-mini (22.7M, fine-tuned) | 0.865 | 0.954 | 0.914 | 1.322 | 7.4 |
| 3 | BM25 (keyword) | 0.856 | 0.940 | 0.905 | 1.407 | — |
| 4 | gte-reranker-modernbert-base (off the shelf) |
0.833 | 0.926 | 0.889 | 1.464 | 53.1 |
| 5 | Hybrid RRF (k = 60) | 0.728 | 0.793 | 0.788 | 2.371 | — |
| 6 | Dense (text-embedding-3-small) |
0.681 | 0.749 | 0.745 | 2.859 | — |
| 7 | ms-marco-MiniLM-L-6-v2 (off the shelf) |
0.643 | 0.817 | 0.750 | 2.210 | 9.4 |
These are reranking numbers — picking the right paper out of nine confusable ones — not open-corpus retrieval.
Details
| Property | |
|---|---|
| Corpus | 44B — arXiv, OpenAlex and Semantic Scholar records unified with OpenReview reviews; ICLR 2024–2025, NeurIPS 2024 |
| Query | an official review's summary, or a meta-review's metareview; ≥ 200 characters |
| Split | by paper — sha256("44b-training-pairs-2026-07-31:" + paper) → [0,1), train < 0.8 ≤ val < 0.9 ≤ test |
| Disjointness | ` |
| Negatives | 8 per query: semantic (HNSW over text-embedding-3-small, ef_search 160) and lexical (Postgres FTS); false-negative screen on Semantic Scholar ID and near-identical title |
| Truncation at scoring | query 600 characters, document 900 characters |
| Frozen | 2026-07-31 |
How to use
import json
rows = [json.loads(l) for l in open("test.jsonl")]
for r in rows:
query_url = r["note_url"] # fetch the review's summary here
candidates = [r["positive"], *r["negatives"]] # fetch title + abstract at each ["url"]
# score all 9 with your reranker; the positive is at index 0 before you shuffle
Licensing
Review text belongs to its authors on OpenReview; abstracts come from Semantic Scholar, OpenAlex and arXiv under their own terms. None of that text is redistributed here. This release ships identifiers, the benchmark construction, and our own measurements — released CC-BY-4.0. A third party resolves the IDs against the same public sources we did, which is the standard MS MARCO / BEIR shape.
📬 Contact
Questions, results, or a model to add to the table? Open a discussion in the Community tab.
Citation
@misc{44beval2026,
title = {44b Eval: a peer-review-queried hard-negative reranking benchmark over ML papers},
author = {NYSgpt},
year = {2026},
url = {https://huggingface.co/datasets/NYSgpt/44b-eval}
}