NanoMTEB-Dutch is a compact Dutch retrieval benchmark containing the MTEB-NL retrieval-family splits. It combines Dutch BEIR-style tasks, legal and public-domain QA, news, tender, web FAQ, Wikipedia, and cross-lingual Belebele retrieval splits.

Usage

from datasets import load_dataset

dataset_id = "hakari-bench/NanoMTEB-Dutch"
split = "argu_ana_nl"

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
argu_ana_nl	199	8624	199	1316.9	1106.0	1500.5	1141.1	1029.0	1399.0
b_bsardnl	200	10000	923	93.8	89.0	117.0	863.2	471.0	964.2
belebele_eng_latn_nld_latn	200	488	200	69.4	68.5	81.5	475.5	456.5	585.2
belebele_nld_latn_eng_latn	200	488	200	81.3	81.0	96.0	529.1	520.0	658.8
belebele_nld_latn_nld_latn	200	488	200	69.4	68.5	81.5	529.1	520.0	658.8
cqadupstack_android	200	10000	200	59.1	56.0	72.0	638.1	505.0	764.0
cqadupstack_english	200	10000	200	49.6	47.0	61.2	521.7	408.0	634.2
cqadupstack_gis	200	10000	200	62.7	58.0	75.0	1036.1	718.0	1206.0
cqadupstack_mathematica	200	10000	200	55.3	53.0	64.2	1166.7	874.0	1461.2
cqadupstack_physics	200	10000	200	62.1	55.0	78.0	870.4	668.0	1093.0
cqadupstack_programmers	200	10000	200	61.2	59.0	76.0	1142.3	893.5	1436.0
cqadupstack_stats	200	10000	200	64.3	60.5	80.0	1097.6	862.0	1373.0
cqadupstack_tex	200	10000	200	53.5	49.5	66.2	1211.8	864.0	1490.2
cqadupstack_webmasters	200	10000	200	58.8	56.5	76.2	761.2	613.0	924.0
cqadupstack_wordpress	200	10000	200	56.5	55.5	66.0	1183.4	811.5	1391.0
dutch_news_articles	200	10000	200	49.0	47.0	60.0	1146.7	799.0	1347.2
fever	200	10000	233	54.9	50.5	64.2	445.7	295.0	539.0
legal_qanl	102	10000	157	104.3	87.0	121.0	665.0	634.0	1061.0
nfcorpus_nl	199	3593	5880	18.5	12.0	22.5	1743.7	1769.0	2052.0
nq	200	10000	242	52.7	49.0	59.0	595.4	530.0	802.0
open_tender	199	10000	199	62.2	52.0	78.0	442.0	338.0	641.0
quora	200	10000	573	51.8	49.0	62.2	66.6	57.0	79.0
sci_fact_nl	200	5183	226	100.1	92.5	121.0	1640.3	1555.0	1978.0
scidocs_nl	200	10000	986	77.7	74.0	91.0	1331.6	1235.0	1581.0
vabb	200	9123	200	74.5	71.5	97.2	837.9	661.0	1167.5
web_faq_nld	200	10000	200	50.4	45.5	57.0	322.2	263.0	378.0
wikipedia_multilingual_nl	200	10000	200	63.5	62.0	76.0	381.0	309.5	496.2

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	-	46.73	55.95	52.40	68.97	78.93	81.54	-	756
argu_ana_nl	regex	29.70	37.23	35.29	92.46	97.49	99.50	100-101	1
b_bsardnl	regex	12.49	27.49	22.34	38.04	61.84	60.33	100-101	42
belebele_eng_latn_nld_latn	regex	47.38	89.18	62.83	61.50	98.50	99.00	100-101	2
belebele_nld_latn_eng_latn	english_porter_stop	32.88	93.06	44.56	44.00	98.50	99.00	100-101	2
belebele_nld_latn_nld_latn	regex	83.64	88.99	89.99	97.00	96.50	99.00	100-101	2
cqadupstack_android	regex	29.44	38.62	38.36	63.00	77.50	78.00	100-101	44
cqadupstack_english	regex	27.69	35.87	32.48	49.50	65.00	68.50	100-101	63
cqadupstack_gis	regex	27.90	32.02	32.72	60.00	68.50	72.00	100-101	56
cqadupstack_mathematica	regex	18.26	19.92	21.81	43.50	52.50	57.50	100-101	85
cqadupstack_physics	regex	32.69	40.20	37.56	62.50	75.00	75.50	100-101	49
cqadupstack_programmers	regex	29.91	39.06	36.38	60.50	76.00	76.00	100-101	48
cqadupstack_stats	regex	28.27	32.24	33.37	57.00	65.50	68.00	100-101	64
cqadupstack_tex	regex	21.06	26.11	29.26	56.50	64.00	67.50	100-101	65
cqadupstack_webmasters	regex	23.07	29.47	29.68	55.50	67.00	72.50	100-101	55
cqadupstack_wordpress	regex	26.08	30.57	33.71	59.50	68.50	72.50	100-101	55
dutch_news_articles	regex	88.68	89.96	89.54	97.50	94.50	99.50	100-101	1
fever	regex	92.21	92.07	92.15	98.83	93.92	99.58	100	0
legal_qanl	regex	81.43	80.50	84.55	97.55	94.16	98.53	100-101	1
nfcorpus_nl	regex	26.83	25.90	26.56	19.99	26.68	26.14	100-101	37
nq	regex	45.05	63.35	54.73	87.75	91.00	98.75	100-101	2
open_tender	regex	67.12	60.44	65.56	80.90	80.90	85.43	100-101	29
quora	regex	83.91	92.89	87.72	94.51	95.92	99.32	100	0
sci_fact_nl	regex	61.60	67.58	67.09	84.53	92.83	95.00	100-101	10
scidocs_nl	regex	13.35	22.64	18.35	27.07	45.84	43.02	100-101	26
vabb	regex	69.52	78.04	75.40	86.50	90.00	92.00	100-101	16
web_faq_nld	regex	76.98	87.76	84.42	90.50	95.50	100.00	100	0
wikipedia_multilingual_nl	regex	84.44	89.48	88.40	96.00	97.50	99.50	100-101	1

Hybrid Safeguard Summary

Safeguard positives: 756
Rows limited by corpus size: 0
Metadata file: reranking_hybrid_metadata.json

Source Links

License

NanoMTEB-Dutch 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: 431

Total file size:

183 MB