NanoMMTEB-v2 is a compact multilingual and multi-domain retrieval benchmark assembled from MMTEB/MTEB retrieval tasks. It includes legal, QA, long-context, dialogue, social, code/community, and reasoning-oriented retrieval splits in the Nano query-corpus-qrels format.

Usage

from datasets import load_dataset

dataset_id = "hakari-bench/NanoMMTEB-v2"
split = "ailastatutes"

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
ailastatutes	50	82	217	3038.4	3063.0	3867.0	1972.6	898.0	1976.0
argu_ana	199	8626	199	1199.8	1002.0	1359.5	1029.6	927.0	1263.8
belebele	376	10000	376	95.4	103.0	110.0	509.2	476.0	606.0
covid	200	10000	204	25.7	24.0	30.0	409.3	275.0	329.0
hagrid	200	493	200	38.4	36.0	46.0	229.6	133.0	327.0
legal_bench_corporate_lobbying	200	319	200	179.7	155.0	225.5	1157.2	1033.0	1490.0
lembpasskey	100	100	100	37.8	38.0	39.0	28060.9	7244.0	29076.0
miracl	200	10000	444	37.2	35.5	45.0	448.2	359.0	596.0
mlqa	196	10000	196	47.4	42.5	57.0	731.3	515.0	899.0
scidocs	200	10000	986	69.8	67.0	80.2	1202.7	1102.0	1415.0
spart_qa	200	1592	384	654.9	621.0	727.5	49.8	55.0	62.0
stack_overflow_qa	200	10000	200	1361.8	908.5	1621.2	1218.1	730.0	1432.0
statcan_dialogue_dataset	200	10000	313	794.8	539.0	895.2	7237.7	2211.5	5829.0
temp_reason_l1	200	10000	200	49.9	51.0	52.0	9.0	9.0	9.0
treccovid	50	10000	4527	69.2	64.5	76.8	1321.6	1359.0	1806.0
twitter_hjerne	77	262	262	165.8	148.0	230.0	128.8	119.0	178.0
wikipedia_multilingual	200	10000	200	59.2	55.0	71.0	383.3	301.0	493.0
wino_grande	200	5095	200	112.0	110.0	122.2	7.7	7.0	9.0

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	-	45.50	53.06	50.71	67.65	80.15	78.83	-	564
ailastatutes	english_porter_stop	20.70	27.25	25.57	100.00	100.00	100.00	82	0
argu_ana	english_porter_stop	34.64	39.98	37.16	95.48	94.97	98.99	100-101	2
belebele	english_porter_stop	9.03	27.81	17.82	22.07	47.87	41.22	100-101	221
covid	wordseg@zh	78.88	75.92	78.73	96.00	93.50	99.00	100-101	2
hagrid	english_porter_stop	98.14	95.70	96.39	99.50	98.00	100.00	100	0
legal_bench_corporate_lobbying	english_porter_stop	89.55	91.10	90.80	100.00	98.00	100.00	100	0
lembpasskey	english_porter_stop	99.63	84.63	85.25	100.00	100.00	100.00	100	0
miracl	stemmer@arabic	57.60	77.75	69.42	89.10	94.54	99.35	100	0
mlqa	wordseg@vi	3.90	9.59	5.34	14.29	55.61	42.35	100-101	113
scidocs	english_porter_stop	20.67	27.73	25.90	42.32	57.59	53.49	100-101	11
spart_qa	english_porter_stop	18.48	25.91	33.82	60.67	54.83	62.33	100-101	37
stack_overflow_qa	english_porter_stop	79.70	88.86	84.57	92.50	94.50	99.00	100-101	2
statcan_dialogue_dataset	stemmer@french	1.12	27.31	15.64	16.58	77.38	72.03	100-101	47
temp_reason_l1	english_porter_stop	1.61	4.88	1.34	3.50	64.50	36.50	100-101	127
treccovid	english_porter_stop	36.27	42.66	45.05	24.61	26.70	29.28	100	0
twitter_hjerne	regex	23.95	62.43	44.02	62.51	89.72	85.35	100-101	2
wikipedia_multilingual	regex	94.25	96.24	94.52	98.50	97.00	100.00	100	0
wino_grande	english_porter_stop	50.84	49.40	61.39	100.00	98.00	100.00	100	0

Hybrid Safeguard Summary

Safeguard positives: 564
Rows limited by corpus size: 50
Metadata file: reranking_hybrid_metadata.json

Source Links

License

NanoMMTEB-v2 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: 370

Total file size:

169 MB