🚧 This dataset is currently under construction. Specifications may change.

NanoMIRACL (with bm25 subset)

A lightweight, evaluation-ready subset of MIRACL, designed for fast multilingual IR benchmarking.

What this dataset is

• A multilingual IR benchmark with 18 language splits: ar, bn, de, en, es, fa, fi, fr, hi, id, ja, ko, ru, sw, te, th, yo, zh.
• Each split contains 50 queries and a corpus of 10,000 documents.
• Built with deterministic sampling for reproducible benchmarking.

Data structure

• corpus: _id, text
• queries: _id, text
• qrels: query-id, corpus-id, score
• bm25: query-id, corpus-ids (A bm25 subset is included for first-stage retrieval and reranking experiments.)

Performance Comparison Across Models

Model	avg	ar	bn	de	en	es	fa	fi	fr	hi	id	ja	ko	ru	sw	te	th	yo	zh
bm25	0.5138	0.5363	0.4760	0.4724	0.5831	0.5582	0.4837	0.6503	0.4241	0.5988	0.5178	0.5569	0.4627	0.4829	0.4472	0.5884	0.5798	0.4664	0.3632
e5-small	0.7181	0.7463	0.7127	0.6579	0.6968	0.7278	0.7203	0.7808	0.6339	0.6612	0.6261	0.7044	0.6730	0.7271	0.7025	0.9779	0.8896	0.6130	0.6737
e5-large	0.7791	0.8186	0.7920	0.7925	0.7386	0.8026	0.7522	0.8347	0.6710	0.7586	0.6229	0.7449	0.6907	0.7927	0.7540	0.9779	0.9174	0.8168	0.7465
bge-m3	0.7873	0.8193	0.8417	0.7337	0.7406	0.7419	0.7733	0.7977	0.7355	0.7402	0.6636	0.7975	0.7266	0.7993	0.7635	0.9926	0.9090	0.8547	0.7402

Subset names

• Split names: ar, bn, de, en, es, fa, fi, fr, hi, id, ja, ko, ru, sw, te, th, yo, zh
• Config names: corpus, queries, qrels, bm25

BM25 tokenization strategy

• bm25 was generated with --auto-select-best-splitter.
• Candidate strategies: transformer, stemmer, wordseg, nltk_stem, nltk_stem_stop, english_regex, english_porter, english_porter_stop, whitespace.
• Selection metric: best nDCG@100 per split.

Selected strategy by split:

Split	Strategy	Details
ar	nltk_stem	NLTK stemmer
bn	whitespace	str.split()
de	nltk_stem_stop	NLTK stemmer + stopword removal
en	english_regex	regex tokenization (no stemming)
es	nltk_stem	NLTK stemmer
fa	whitespace	str.split()
fi	nltk_stem_stop	NLTK stemmer + stopword removal
fr	nltk_stem_stop	NLTK stemmer + stopword removal
hi	stemmer	PyStemmer (hindi)
id	transformer	tokenizer: Qwen/Qwen3-0.6B
ja	wordseg	ja (fugashi + unidic-lite)
ko	transformer	tokenizer: Qwen/Qwen3-0.6B
ru	nltk_stem	NLTK stemmer
sw	whitespace	str.split()
te	whitespace	str.split()
th	wordseg	th (pythainlp newmm)
yo	transformer	tokenizer: Qwen/Qwen3-0.6B
zh	transformer	tokenizer: Qwen/Qwen3-0.6B

Upstream source

• Original dataset: MIRACL
• Base processed source: hotchpotch/miracl-hf-unified

License

Other. This dataset is derived from MIRACL and follows upstream licensing and attribution requirements.