BRIGHT indexes
Ready-to-use PyTerrier indexes for all 12
BRIGHT subsets. Document ids are those of
irds:bright/<subset>, so retrieved docnos match the BRIGHT qrels
directly; the one exception is PISA on stackoverflow (see Notes).
| Folder | What it is |
|---|---|
pisa/ |
PISA inverted index (pyterrier-pisa), Porter2 stemming, Terrier stopwords; BM25 data precomputed for k1=0.9, b=0.4 |
terrier/ |
Terrier inverted + direct index, Porter stemming, Terrier stopwords; stores document text (meta) for re-rankers |
qwen3-embedding-0.6b/ |
pyterrier-dr FlexIndex of Qwen/Qwen3-Embedding-0.6B document vectors (1024-dim, L2-normalised) |
Each folder <type>/<subset>/ holds one PyTerrier artifact (artifact.tar.lz4) and its own README with
usage and build details. PISA covers every subset; Terrier and Qwen indexes exist for some:
| Subset | pisa/ |
terrier/ |
qwen3-embedding-0.6b/ |
|---|---|---|---|
| biology | β | β | β |
| earth-science | β | β | β |
| economics | β | β | β |
| psychology | β | β | β |
| robotics | β | β | β |
| stackoverflow | β | β | β |
| sustainable-living | β | β | β |
| leetcode | β | β | β |
| pony | β | β | β |
| aops | β | β | β |
| theoremqa-theorems | β | β | β |
| theoremqa-questions | β | β | β |
Indexes
BM25 scores are nDCG@10 / R@100 with k1=0.9, b=0.4 on all queries of the subset (original text field).
Terrier's default BM25 parameters (k1=1.2, b=0.75) score much lower on BRIGHT; use k1=0.9, b=0.4.
| Index | Docs | Size | BM25 nDCG@10 / R@100 |
|---|---|---|---|
pisa/biology |
57,359 | 22 MB | 0.170 / 0.468 |
pisa/earth-science |
121,249 | 49 MB | 0.256 / 0.584 |
pisa/economics |
50,220 | 19 MB | 0.140 / 0.402 |
pisa/psychology |
52,835 | 21 MB | 0.127 / 0.402 |
pisa/robotics |
61,961 | 16 MB | 0.149 / 0.453 |
pisa/stackoverflow |
107,081 | 62 MB | 0.174 / 0.462 |
pisa/sustainable-living |
60,792 | 22 MB | 0.126 / 0.442 |
pisa/leetcode |
413,932 | 292 MB | 0.242 / 0.480 |
pisa/pony |
7,894 | 2 MB | 0.088 / 0.259 |
pisa/aops |
188,002 | 119 MB | 0.064 / 0.224 |
pisa/theoremqa-theorems |
23,839 | 13 MB | 0.049 / 0.167 |
pisa/theoremqa-questions |
188,002 | 119 MB | 0.096 / 0.238 |
terrier/biology |
57,359 | 26 MB | 0.188 / 0.453 |
terrier/earth-science |
121,249 | 55 MB | 0.301 / 0.610 |
terrier/robotics |
61,961 | 20 MB | 0.156 / 0.450 |
qwen3-embedding-0.6b/biology |
57,359 | 235 MB | β |
qwen3-embedding-0.6b/earth-science |
121,249 | 499 MB | β |
qwen3-embedding-0.6b/economics |
50,220 | 206 MB | β |
qwen3-embedding-0.6b/psychology |
52,835 | 217 MB | β |
qwen3-embedding-0.6b/robotics |
61,961 | 252 MB | β |
qwen3-embedding-0.6b/pony |
7,894 | 32 MB | β |
Notes
- stackoverflow docnos (PISA). PISA cannot store docnos with spaces; 8 stackoverflow docnos had each
space replaced by
_.pisa/stackoverflowshipsdocnos_with_space.json(sanitised β original); map the run through it before evaluating:run['docno'] = run['docno'].replace(json.load(open(os.path.join(index.path, 'docnos_with_space.json')))). No other subset needs this. - Excluded documents (aops, leetcode, theoremqa-questions). BRIGHT marks some documents per query as
excluded (qrels label
-100). Following BRIGHT's official evaluation, the scores above remove them from each query's ranking before scoring; plainpt.Experimentwould leave them in place and take top-k slots.
Usage
pip install python-terrier pyterrier-pisa pyterrier-dr
import pyterrier as pt
BASE = 'https://huggingface.co/datasets/mandeep-rathee/bright-indexes/resolve/main'
def bright_index(kind, subset): # kind: 'pisa' | 'terrier' | 'qwen3-embedding-0.6b'
return pt.Artifact.from_url(f'{BASE}/{kind}/{subset}/artifact.tar.lz4')
topics = pt.get_dataset('irds:bright/biology').get_topics('text')
qrels = pt.get_dataset('irds:bright/biology').get_qrels()
# PISA BM25
pisa = bright_index('pisa', 'biology')
bm25 = pt.rewrite.tokenise() >> pisa.bm25(k1=0.9, b=0.4, num_results=1000) >> pt.rewrite.reset()
# Terrier BM25
terrier = bright_index('terrier', 'biology')
tr_bm25 = (pt.rewrite.tokenise()
>> pt.terrier.Retriever(terrier, wmodel='BM25', num_results=1000,
controls={'bm25.k_1': 0.9, 'bm25.b': 0.4})
>> pt.rewrite.reset())
from ir_measures import nDCG, R
pt.Experiment([bm25, tr_bm25], topics, qrels, [nDCG@10, R@100], names=['pisa', 'terrier'])
The Qwen FlexIndex needs queries encoded with the same model. Documents were encoded without an
instruction; queries used the Qwen3-Embedding format
Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery: <query>.
See qwen3-embedding-0.6b/biology/README.md
for a full sentence-transformers example.
License and citation
Derived from the BRIGHT corpus and released under the same license (CC BY 4.0). Please cite BRIGHT:
@inproceedings{su2025bright,
title = {{BRIGHT}: A Realistic and Challenging Benchmark for Reasoning-Intensive Retrieval},
author = {Su, Hongjin and Yen, Howard and Xia, Mengzhou and Shi, Weijia and Muennighoff, Niklas and Wang, Han-yu and Liu, Haisu and Shi, Quan and Siegel, Zachary S. and Tang, Michael and Sun, Ruoxi and Yoon, Jinsung and Arik, Sercan O. and Chen, Danqi and Yu, Tao},
booktitle = {International Conference on Learning Representations (ICLR)},
year = {2025}
}
- Downloads last month
- 42