NanoMTEB-Polish is a compact Polish retrieval benchmark assembled from Polish MTEB-family retrieval tasks. It includes Polish CQADupStack domains, FiQA, Natural Questions, PUGG information retrieval, and Quora-style duplicate-question retrieval.

Usage

from datasets import load_dataset

dataset_id = "hakari-bench/NanoMTEB-Polish"
split = "cqadupstack_android"

queries = load_dataset(dataset_id, "queries", split=split)
corpus = load_dataset(dataset_id, "corpus", split=split)
qrels = load_dataset(dataset_id, "qrels", split=split)
reranking_candidates = load_dataset(dataset_id, "reranking_hybrid", split=split)

Data Layout

This dataset uses six Hugging Face Datasets configs:

corpus: documents with _id and text
queries: queries with _id and text
qrels: positive relevance labels with query-id and corpus-id
bm25: BM25 candidate lists with query-id and corpus-ids
harrier_oss_v1_270m: dense candidate lists from microsoft/harrier-oss-v1-270m
reranking_hybrid: RRF candidate lists built from bm25 and harrier_oss_v1_270m

Each config has the same Nano split names.

Candidate Construction

bm25: local BM25 top-500 with automatic language-aware tokenization. The resolved tokenizer is shown in the Candidate Quality table, for example wordseg@ja.
harrier_oss_v1_270m: dense top-500 from microsoft/harrier-oss-v1-270m. In tables this is shown as Dense; Dense means microsoft/harrier-oss-v1-270m with the web_search_query prompt for queries and cosine similarity over normalized embeddings.
reranking_hybrid: RRF over bm25 and harrier_oss_v1_270m using rrf_k=100, keeping the RRF top-100.

Safeguard means rank 101 is appended only when RRF top-100 contains no qrels-positive document.

Split Statistics

Length statistics are character counts computed with len(str(text)).

Nano split	Queries	Corpus	Qrels	Query chars avg	Query chars p50	Query chars p75	Doc chars avg	Doc chars p50	Doc chars p75
cqadupstack_android	200	10000	809	59.3	56.5	72.0	626.7	484.0	733.0
cqadupstack_english	200	10000	1356	46.5	43.0	58.0	488.1	383.0	596.0
cqadupstack_gis	200	10000	313	60.6	58.0	74.0	966.0	667.0	1098.0
cqadupstack_mathematica	200	10000	506	50.4	46.0	62.2	1088.5	782.0	1304.0
cqadupstack_physics	200	10000	621	58.8	51.5	74.0	814.7	622.0	1014.0
cqadupstack_programmers	200	10000	634	59.1	57.0	71.2	1075.3	843.0	1343.0
cqadupstack_stats	200	10000	373	61.0	60.0	76.2	1016.6	792.0	1253.0
cqadupstack_tex	200	10000	843	50.3	46.0	60.2	1106.1	753.0	1285.0
cqadupstack_webmasters	200	10000	882	59.8	59.0	74.2	739.2	592.0	886.2
cqadupstack_wordpress	200	10000	344	55.8	54.0	66.0	1040.6	736.5	1220.2
fiqa	200	10000	534	68.5	66.5	83.0	808.8	551.0	991.0
nq	200	10000	251	48.6	46.0	56.0	616.8	528.0	831.0
pugg	200	10000	200	36.2	33.0	43.0	850.3	861.0	902.0
quora	200	10000	485	52.5	47.0	62.0	64.6	55.0	77.0

Candidate Quality

nDCG@10 and Recall@100 are computed from the included candidate rankings against the included qrels, then reported as 0-100 scores such as 52.45. Recall@100 uses only the top 100 candidates; an optional rank-101 safeguard positive is not counted in Recall@100.

Dense means microsoft/harrier-oss-v1-270m with the web_search_query prompt and cosine similarity.

Nano split	BM25 tokenizer	BM25 nDCG@10	Dense nDCG@10	Hybrid nDCG@10	BM25 Recall@100	Dense Recall@100	Hybrid Recall@100	Hybrid candidates	Safeguard positives
Mean	-	34.24	42.71	40.88	60.23	69.78	73.42	-	505
cqadupstack_android	regex	33.79	41.39	41.21	60.67	71.38	74.98	100-101	25
cqadupstack_english	regex	31.88	39.26	37.25	46.63	61.06	61.90	100-101	45
cqadupstack_gis	regex	24.23	28.61	31.43	56.21	60.21	67.65	100-101	56
cqadupstack_mathematica	regex	21.29	21.71	24.11	39.69	48.61	50.69	100-101	75
cqadupstack_physics	regex	33.59	43.06	40.24	60.11	76.83	75.88	100-101	27
cqadupstack_programmers	regex	31.91	32.75	36.07	58.86	64.23	70.23	100-101	41
cqadupstack_stats	regex	26.62	33.75	33.14	52.86	60.71	64.13	100-101	55
cqadupstack_tex	regex	25.55	28.05	31.47	51.62	56.92	61.82	100-101	50
cqadupstack_webmasters	regex	24.40	30.45	31.62	49.85	63.50	66.71	100-101	42
cqadupstack_wordpress	regex	31.39	29.51	32.89	57.79	60.76	69.66	100-101	50
fiqa	regex	23.53	38.90	35.74	52.41	70.98	72.20	100-101	26
nq	regex	30.26	61.54	43.63	76.92	92.58	94.75	100-101	9
pugg	regex	63.90	78.17	71.46	87.50	93.00	98.00	100-101	4
quora	regex	77.04	90.73	82.07	92.04	96.13	99.25	100	0

Hybrid Safeguard Summary

Safeguard positives: 505
Rows limited by corpus size: 0
Metadata file: reranking_hybrid_metadata.json

Source Links

License

NanoMTEB-Polish is a derived dataset. Users must comply with the licenses, terms, and attribution requirements of the upstream MTEB task sources and their original datasets.

Downloads last month: 269

Total file size:

103 MB