bm25-benchmarks 0.0.3

CLI tools for running BM25 benchmark evals.

BM25 Benchmarks

CLI

Installation

From PyPI with pip:

pip install bm25-benchmarks
bm25-benchmark install bm25s

From GitHub with pip:

pip install "bm25-benchmarks @ git+https://github.com/xhluca/bm25-benchmarks.git"
bm25-benchmark install bm25s

With uv as a globally available tool:

uv tool install bm25-benchmarks
bm25-benchmark install bm25s --installer uv

With uv into the current virtual environment:

uv pip install "bm25-benchmarks @ git+https://github.com/xhluca/bm25-benchmarks.git"
bm25-benchmark install bm25s --installer uv

For local development:

pip install -e ".[bm25s]"
# or
uv pip install -e ".[bm25s]"

Install a different backend by replacing bm25s with rank, bm25-pt, pyserini, elastic, pisa, or all.

For rank-bm25, use the CLI installer so the pinned Git dependency is installed for you. PyPI rejects direct Git dependencies in package metadata, so the pin is kept in requirements-rank-bm25.txt and the CLI installer:

bm25-benchmark install rank
bm25-benchmark install all
bm25-benchmark install rank --installer uv

Usage

The benchmark package exposes one CLI for running evals:

bm25-benchmark --help
bm25-benchmark models
bm25-benchmark datasets

Run an eval by choosing a backend and dataset:

bm25-benchmark eval bm25s -d fiqa
bm25-benchmark eval rank-bm25 -d fiqa --samples 1000
bm25-benchmark eval pyserini -d fiqa --threads 4
bm25-benchmark eval elastic -d fiqa --hostname localhost
bm25-benchmark eval pisa -d fiqa
bm25-benchmark eval bm25-pt -d fiqa --batch-size 32

Common eval options include:

bm25-benchmark eval bm25s -d fiqa -d scifact --result-dir results --save-dir datasets
bm25-benchmark eval bm25s -d fiqa,scifact --num-runs 3
bm25-benchmark eval bm25s -d fiqa --dry-run

Use bm25-benchmark eval <backend> --help to see backend-specific options.

The module form is also available:

python -m benchmark eval bm25s -d fiqa

Benchmarking

The legacy module entry points are still available:

# For bm25_pt
python -m benchmark.on_bm25_pt -d "<dataset>"

# For rank-bm25
python -m benchmark.on_rank_bm25 -d "<dataset>"

# for Pyserini
python -m benchmark.on_pyserini -d "<dataset>"

# For elastic, After starting the server, run:
python -m benchmark.on_elastic -d "<dataset>"

# for PISA
python -m benchmark.on_pisa -d "<dataset>"

# For bm25s
python -m benchmark.on_bm25s -d "<dataset>"

where <dataset> is the name of the dataset to be used.

BM25S options

For bm25s, you can specify which scoring methods and retrieval backends to benchmark:

# Default: runs the jit scorer and numba backend
bm25-benchmark eval bm25s -d fiqa

# Specify scorers (uncompiled, legacy, jit)
bm25-benchmark eval bm25s -d fiqa --scorers legacy jit
bm25-benchmark eval bm25s -d fiqa --scorers jit
bm25-benchmark eval bm25s -d fiqa --scorers uncompiled legacy jit

# Specify backends (jax, numba, numpy)
bm25-benchmark eval bm25s -d fiqa --backends numba
bm25-benchmark eval bm25s -d fiqa --backends jax numba numpy

# Combine both
bm25-benchmark eval bm25s -d fiqa --scorers jit --backends numba

Scorer options:

• uncompiled: Default NumPy implementation (optimized with np.add.at)
• legacy: Legacy implementation (similar to uncompiled, kept for comparison)
• jit: Numba JIT-compiled version (fastest after warmup)

Backend options:

• jax: JAX-based retrieval
• numba: Numba JIT-compiled retrieval (default)
• numpy: Pure NumPy retrieval

Available datasets

The available datasets are public BEIR datasets: trec-covid, nfcorpus, fiqa, arguana, webis-touche2020, quora, scidocs, scifact, cqadupstack, nq, msmarco, hotpotqa, dbpedia-entity, fever, climate-fever,

Sampling during benchmarking

For rank-bm25, due to the long runtime, we can sample queries

bm25-benchmark eval rank-bm25 -d "<dataset>" --samples <num_samples>

Rank-bm25 variants

For rank-bm25, we can also specify the method with --method to be used:

• rank (default)
• bm25l
• bm25+

Results will be saved in results/ directory.

Elasticsearch server

If you want to use elastic search, you need to start the server first.

First, download the elastic search from here. You will get a file, e.g. elasticsearch-8.14.0-linux-x86_64.tar.gz. Extract the file and ensure it is in the same directory as the bm25-benchmarks directory.

wget https://artifacts.elastic.co/downloads/elasticsearch/elasticsearch-8.14.0-linux-x86_64.tar.gz
tar -xzf elasticsearch-8.14.0-linux-x86_64.tar.gz
# remove the tar file
rm elasticsearch-8.14.0-linux-x86_64.tar.gz

Then, start the server with the following command:

./elasticsearch-8.14.0/bin/elasticsearch -E xpack.security.enabled=false -E thread_pool.search.size=1 -E thread_pool.write.size=1

Results

The results are benchmarked using Kaggle notebooks to ensure reproducibility. Each one is run on single-core, Intel Xeon CPU @ 2.20GHz, using 30GB RAM.

The shorthands used are:

• BM25PT for bm25_pt
• PSRN for pyserini
• R-BM25 for rank-bm25
• BM25S for bm25, and BM25S+J for Numba JIT version of bm25s (v0.2.0+)
• ES for elasticsearch
• PISA for the Pisa Engine (via the pyterrier_pisa Python bindings)
• OOM for out-of-memory error
• DNT for did not terminate (i.e. went over 12 hours)

Queries per second

dataset	PISA	BM25S+J	BM25S	ES	PSRN	PT	R-BM25
arguana	270.53	869.95	573.91	13.67	11.95	110.51	2
climate-fever	35.95	38.49	13.09	4.02	8.06	OOM	0.03
cqadupstack	362.39	396.5	170.91	13.38	DNT	OOM	0.77
dbpedia-entity	197.45	71.8	13.44	10.68	12.69	OOM	0.11
fever	81.42	53.84	20.19	7.45	10.52	OOM	0.06
fiqa	714.35	1237.39	717.78	16.96	12.51	20.52	4.46
hotpotqa	54.98	47.16	20.88	7.11	10.41	OOM	0.04
msmarco	178.65	39.18	12.2	11.88	11.01	OOM	0.07
nfcorpus	5111.72	5696.21	1196.16	45.84	32.94	256.67	224.66
nq	168.12	109.47	41.85	12.16	11.04	OOM	0.1
quora	735.20	479.71	272.04	21.8	15.58	6.49	1.18
scidocs	818.97	1448.32	767.05	17.93	14.1	41.34	9.01
scifact	1463.73	2787.84	1317.12	20.81	15.02	184.3	47.6
trec-covid	282.94	483.84	85.64	7.34	8.53	3.73	1.48
webis-touche2020	431.12	390.03	60.59	13.53	12.36	OOM	1.1

Notes:

• For Rank-BM25, larger datasets are ran with 1000 samples rather than the full dataset to ensure it finishes within 12h (limit for Kaggle notebooks).
• For ES and BM25S, we can set a number of threads to use. However, you might not see an improvement, in fact you might even see a decrease in throughput in the case of BM25S due to how multi-threading is implemented. Click below to see the results.

Show BM25S & ES multi-threaded (4T) performance (Q/s)

dataset	PISA	BM25S	ES
arguana	590.93	211	33.37
climate-fever	91.68	22.06	8.13
cqadupstack	945.66	248.87	27.76
dbpedia-entity	478.26	26.18	15.49
fever	222.08	47.03	14.07
fiqa	1382.32	449.82	36.33
hotpotqa	134.60	45.02	10.35
msmarco	393.16	21.64	18.19
nfcorpus	6706.53	784.24	81.07
nq	423.54	77.49	19.18
quora	1892.98	308.58	43.02
scidocs	1757.44	614.23	46.36
scifact	2480.86	645.88	50.93
trec-covid	676.40	100.88	13.5
webis-touche2020	938.57	202.39	26.55

Show normalized table wrt Rank-BM25

dataset	PISA	BM25S	ES	PSRN	PT	Rank
arguana	135.27	286.96	6.84	5.98	55.26	1
climate-fever	1198.33	436.33	134	268.67	nan	1
cqadupstack	470.64	221.96	17.38	nan	nan	1
dbpedia-entity	1795.00	122.18	97.09	115.36	nan	1
fever	1357.00	336.5	124.17	175.33	nan	1
fiqa	160.17	160.94	3.8	2.8	4.6	1
hotpotqa	1374.50	522	177.75	260.25	nan	1
msmarco	2552.14	174.29	169.71	157.29	nan	1
nfcorpus	22.75	5.32	0.2	0.15	1.14	1
nq	1681.20	418.5	121.6	110.4	nan	1
quora	623.05	230.54	18.47	13.2	5.5	1
scidocs	90.90	85.13	1.99	1.56	4.59	1
scifact	30.75	27.67	0.44	0.32	3.87	1
trec-covid	191.18	57.86	4.96	5.76	2.52	1
webis-touche2020	391.93	55.08	12.3	11.24	nan	1

Stats

	# Docs	# Queries	# Tokens
msmarco	8,841,823	6,980	340,859,891
hotpotqa	5,233,329	7,405	169,530,287
trec-covid	171,332	50	20,231,412
webis-touche2020	382,545	49	74,180,340
arguana	8,674	1,406	947,470
fiqa	57,638	648	5,189,035
nfcorpus	3,633	323	614,081
climate-fever	5,416,593	1,535	318,190,120
nq	2,681,468	3,452	148,249,808
scidocs	25,657	1,000	3,211,248
quora	522,931	10,000	4,202,123
dbpedia-entity	4,635,922	400	162,336,256
cqadupstack	457,199	13,145	44,857,487
fever	5,416,568	6,666	318,184,321
scifact	5,183	300	812,074

Indexing time (docs/s)

The following results follow the same setup as the queries/s benchmarks described above (single-core).

dataset	PISA	BM25S	ES	PSRN	PT	Rank
arguana	3432.50	4314.79	3591.63	1225.18	638.1	5021.3
climate-fever	5462.73	4364.43	3825.89	6880.42	nan	7085.51
cqadupstack	3963.76	4800.89	3725.43	nan	nan	5370.32
dbpedia-entity	9019.62	7576.28	6333.82	8501.7	nan	9110.36
fever	4903.06	4921.88	3879.63	7007.5	nan	5482.64
fiqa	4426.92	5959.25	4035.11	3735.38	421.51	6455.53
hotpotqa	9883.85	7420.39	5455.6	10342.5	nan	9407.9
msmarco	10205.53	7480.71	5391.29	9686.07	nan	12455.9
nfcorpus	2381.11	3169.4	1688.15	692.05	442.2	3579.47
nq	7122.05	6083.86	5742.13	6652.33	nan	6048.85
quora	38512.02	28002.4	8189.75	22818.5	6251.26	47609.2
scidocs	3085.13	4107.46	3008.45	2137.64	312.72	4232.15
scifact	2449.91	3253.63	2649.57	880.53	442.61	3792.84
trec-covid	4642.59	4600.14	2966.98	3768.1	406.37	4672.62
webis-touche2020	2228.10	2971.96	2484.87	2718.41	nan	3115.96

NDCG@10

We use abbreviations for datasets of BEIR benchmarks.

Click to show dataset abbreviations

• AG for arguana
• CD for cqadupstack
• CF for climate-fever
• DB for dbpedia-entity
• FQ for fiqa
• FV for fever
• HP for hotpotqa
• MS for msmarco
• NF for nfcorpus
• NQ for nq
• QR for quora
• SD for scidocs
• SF for scifact
• TC for trec-covid
• WT for webis-touche2020

k1	b	method	Avg.	AG	CD	CF	DB	FQ	FV	HP	MS	NF	NQ	QR	SD	SF	TC	WT
0.9	0.4	Lucene	41.1	40.8	28.2	16.2	31.9	23.8	63.8	62.9	22.8	31.8	30.5	78.7	15.0	67.6	58.9	44.2
1.2	0.75	ATIRE	39.9	48.7	30.1	13.7	30.3	25.3	50.3	58.5	22.6	31.8	29.1	80.5	15.6	68.1	61.0	33.2
1.2	0.75	BM25+	39.9	48.7	30.1	13.7	30.3	25.3	50.3	58.5	22.6	31.8	29.1	80.5	15.6	68.1	61.0	33.2
1.2	0.75	BM25L	39.5	49.6	29.8	13.5	29.4	25.0	46.6	55.9	21.4	32.2	28.1	80.3	15.8	68.7	62.9	33.0
1.2	0.75	Lucene	39.9	48.7	30.1	13.7	30.3	25.3	50.3	58.5	22.6	31.8	29.1	80.5	15.6	68.0	61.0	33.2
1.2	0.75	Robertson	39.9	49.2	29.9	13.7	30.3	25.4	50.3	58.5	22.6	31.9	29.2	80.4	15.5	68.3	59.0	33.8
1.5	0.75	ES	42.0	47.7	29.8	17.8	31.1	25.3	62.0	58.6	22.1	34.4	31.6	80.6	16.3	69.0	68.0	35.4
1.5	0.75	Lucene	39.7	49.3	29.9	13.6	29.9	25.1	48.1	56.9	21.9	32.1	28.5	80.4	15.8	68.7	62.3	33.1
1.5	0.75	PSRN	40.0	48.4	29.8	14.2	30.0	25.3	50.0	57.6	22.1	32.6	28.6	80.6	15.6	68.8	63.4	33.5
1.5	0.75	PT	45.0	44.9	--	--	--	22.5	--	--	--	31.9	--	75.1	14.7	67.8	58.0	--
1.5	0.75	Rank	39.6	49.5	29.6	13.6	29.9	25.3	49.3	58.1	21.1	32.1	28.5	80.3	15.8	68.5	60.1	32.9
1.2	0.75	PISA	38.8	41.1	27.8	13.9	30.5	24.5	49.2	58.2	22.8	34.3	28.2	72.0	15.7	68.9	64.2	30.9

Recall@1000

k1	b	method	Avg.	AG	CD	CF	DB	FQ	FV	HP	MS	NF	NQ	QR	SD	SF	TC	WT
0.9	0.4	Lucene	77.3	98.8	71.1	63.3	67.5	74.3	95.7	88.0	85.3	47.7	89.6	99.5	56.5	97.0	39.2	86.0
1.2	0.75	ATIRE	77.4	99.3	73.0	59.0	67.0	76.5	94.2	86.8	85.7	47.8	89.8	99.5	57.3	97.0	40.3	87.2
1.2	0.75	BM25+	77.4	99.3	73.0	59.0	67.0	76.5	94.2	86.8	85.7	47.8	89.8	99.5	57.3	97.0	40.3	87.2
1.2	0.75	BM25L	77.2	99.4	73.4	57.3	66.1	77.3	93.7	85.7	85.0	47.7	89.3	99.5	57.7	97.0	40.8	87.5
1.2	0.75	Lucene	77.4	99.3	73.0	59.0	67.0	76.5	94.2	86.8	85.6	47.8	89.8	99.5	57.3	97.0	40.3	87.2
1.2	0.75	Robertson	77.4	99.3	73.2	59.1	66.7	76.8	94.2	86.8	85.9	47.5	89.8	99.5	57.3	96.7	40.2	87.4
1.5	0.75	ES	76.9	99.2	74.2	58.8	63.6	76.7	95.9	85.2	85.1	39.0	90.8	99.6	57.9	98.0	41.3	88.0
1.5	0.75	Lucene	77.2	99.3	73.3	57.8	66.3	77.2	93.8	86.1	85.2	47.7	89.5	99.6	57.5	97.0	40.6	87.4
1.5	0.75	PSRN	76.7	99.2	74.2	58.7	66.2	76.7	94.2	86.4	85.1	37.1	89.4	99.6	57.4	97.7	41.1	87.2
1.5	0.75	PT	73.0	98.3	--	--	--	72.5	--	--	--	51.0	--	98.9	56.0	97.8	36.3	--
1.5	0.75	Rank	77.1	99.4	73.4	57.5	66.4	77.4	93.6	87.7	82.6	47.6	89.5	99.5	57.4	96.7	40.5	87.5
1.2	0.75	PISA	77.1	98.7	72.2	60.2	67.7	76.5	93.7	86.8	86.9	38.4	89.1	98.9	56.9	97.0	45.9	87.4

Links

• BM25+
• BM25L
• ATIRE
• Robertson
• Lucene (k1=1.2, b=0.75)
• Lucene (k1=0.9, b=0.4)
• Lucene (k1=1.5, b=0.75): DB, MS, FV, CF, NQ, HP, Remaining
• ES: FV, CF, NQ, MS, HP, DB, Remaining
• PT: MS, CF, FV, DB, HP, NQ, WT, CD, Remaining
• Rank: DB, HP, CF, FV, MS, NQ, CD, Remaining
• PSRN: CD, FV, HP, MS, DB, NQ, Remaining
• PISA: NQ, DB, CF, HP, FV, MS, CD, Remaining
• BM25+J: Sub-1m, remaining