BrowseComp-Plus indexes (PyTerrier)
pisa/document/ is a PISA inverted index of the
BrowseComp-Plus document corpus
(Tevatron/browsecomp-plus-corpus,
revision b27b02bc): 100,195 web documents, indexed with
pyterrier-pisa 0.3.1 defaults (Porter2 stemmer, Terrier
stopwords). Docnos are the corpus docids, so results match the BrowseComp-Plus qrels directly.
Block-max BM25 data for k1=0.9, b=0.4 is included (1.98 GB compressed).
Usage
pip install python-terrier pyterrier-pisa
import pyterrier as pt
index = pt.Artifact.from_url(
'https://huggingface.co/datasets/mandeep-rathee/browsecomp-plus-indexes/resolve/main/pisa/document/artifact.tar.lz4')
bm25 = pt.rewrite.tokenise() >> index.bm25(k1=0.9, b=0.4, num_results=1000) >> pt.rewrite.reset()
run = bm25(topics) # topics: DataFrame with qid, query
The BrowseComp-Plus queries and relevance labels
(Tevatron/browsecomp-plus) are obfuscated
upstream to keep them out of web crawls; de-obfuscate them with the official
decrypt_dataset.py.
Please don't publish the decrypted queries.
Benchmarks
BM25 (k1=0.9, b=0.4) on all 830 queries, top 1000, scored against the gold and evidence documents. Values in %; in brackets the BM25 document-corpus numbers from the BrowseComp-Plus paper (Table 2, Pyserini).
| Qrels | R@5 | R@100 | R@1000 | nDCG@10 |
|---|---|---|---|---|
| gold | 2.3 (1.4) | 9.0 (6.1) | 22.8 (17.3) | 2.3 (1.7) |
| evidence | 1.8 (1.2) | 7.3 (4.7) | 19.6 (13.7) | 2.3 (1.6) |
This index scores about 1.4× above the paper's Lucene/Pyserini BM25 on every metric, presumably due to the different analyzers (stemmer, stopwords, query tokenisation); both are very low, since documents average 5,179 words.
Reproduction
from pyterrier_pisa import PisaIndex
# corpus_iter(): {'docno': docid, 'text': text} for each row of Tevatron/browsecomp-plus-corpus
PisaIndex('browsecomp-plus.document.pisa', threads=32).index(corpus_iter())
Metadata
{
"type": "sparse_index",
"format": "pisa",
"package_hint": "pyterrier-pisa",
"stemmer": "porter2",
"doc_count": 100195
}
License and citation
Derived from the BrowseComp-Plus corpus and released under the same license (MIT). Please cite:
@article{chen2025browsecompplus,
title = {BrowseComp-Plus: A More Fair and Transparent Evaluation Benchmark of Deep-Research Agent},
author = {Chen, Zijian and Ma, Xueguang and Zhuang, Shengyao and Nie, Ping and Zou, Kai and Liu, Andrew and Green, Joshua and Patel, Kshama and Meng, Ruoxi and Su, Mingyi and Sharifymoghaddam, Sahel and Li, Yanxi and Hong, Haoran and Shi, Xinyu and Liu, Xuye and Thakur, Nandan and Zhang, Crystina and Gao, Luyu and Chen, Wenhu and Lin, Jimmy},
journal = {arXiv preprint arXiv:2508.06600},
year = {2025}
}
- Downloads last month
- 23