NanoMIRACL

This dataset is a Nano-style retrieval dataset prepared for evaluation with NanoIR Benchmark.

NanoMIRACL is derived from hotchpotch/miracl-hf-unified. It follows the Hugging Face Datasets layout convention used by sentence-transformers/NanoBEIR-en: each Nano split has separate corpus, queries, and qrels tables, and BM25 candidates are provided separately in a bm25 table. This layout follows the NanoBEIR-style evaluation approach summarized in NanoBEIR.

NanoMIRACL contains multilingual retrieval splits derived from hotchpotch/miracl-hf-unified. The testB query split is used when available, with a fallback recorded in split metadata when the requested split is not usable.

Source Links

• MIRACL unified source dataset

Data Layout

This dataset uses four 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

Each config has the same Nano split names. The exact parquet paths are defined in the dataset card metadata above. If a regenerated dataset uses a different schema, config name, path layout, or field name, revise this section before publishing the README.

Construction Steps

This dataset was built as follows. If the actual generation procedure differs, revise this section before publishing the README.

• For each selected query, one positive document is kept and negatives are sampled from that query's source negatives before corpus fill.

1. Use hotchpotch/miracl-hf-unified as the upstream benchmark or dataset family.
2. Load source datasets from hotchpotch/miracl-hf-unified.
3. Use testB queries when usable, with recorded per-language fallback when necessary.
4. Create one Nano split for each selected source retrieval task.
5. Keep up to 200 eligible queries per Nano split.
6. Include all qrels-positive documents for the selected queries.
7. Fill the corpus from source corpus order up to 10000 documents.
8. Remove exact duplicate document text within each split. If a removed duplicate was referenced by qrels, rewrite qrels to the retained document id.
9. Store corpus text in the generated document text field.
10. Generate BM25 top-100 candidates with per-split auto tokenization, or the per-split tokenizer shown below.
11. If a qrels-positive document is missing from the raw BM25 result, insert it into the final bm25 candidate list by replacing a tail non-positive candidate.

BM25 Subset Policy

The bm25 config is a candidate subset for first-stage retrieval and reranking. It is not a separate source dataset. Each row contains one query id and a ranked list of up to 100 corpus ids.

BM25 candidates are generated from the selected corpus for each split. When a qrels-positive document is not present in the raw BM25 top-100 results, the missing positive is forced into the final candidate list by replacing a tail candidate that is not positive for that query. Candidate ids are kept unique after replacement.

Split Mapping

Nano split	Source task	Source dataset	Queries	Corpus	Qrels
ar	ar_queries	hotchpotch/miracl-hf-unified	200	1854	200
bn	bn_queries	hotchpotch/miracl-hf-unified	200	1731	200
de	de_queries	hotchpotch/miracl-hf-unified	200	1748	200
en	en_queries	hotchpotch/miracl-hf-unified	200	1657	200
es	es_queries	hotchpotch/miracl-hf-unified	200	1312	200
fa	fa_queries	hotchpotch/miracl-hf-unified	200	1858	200
fi	fi_queries	hotchpotch/miracl-hf-unified	200	1828	200
fr	fr_queries	hotchpotch/miracl-hf-unified	200	1777	200
hi	hi_queries	hotchpotch/miracl-hf-unified	200	1748	200
id	id_queries	hotchpotch/miracl-hf-unified	200	1520	200
ja	ja_queries	hotchpotch/miracl-hf-unified	200	1846	200
ko	ko_queries	hotchpotch/miracl-hf-unified	200	2419	200
ru	ru_queries	hotchpotch/miracl-hf-unified	200	1727	200
sw	sw_queries	hotchpotch/miracl-hf-unified	200	1600	200
te	te_queries	hotchpotch/miracl-hf-unified	84	754	84
th	th_queries	hotchpotch/miracl-hf-unified	200	1897	200
yo	yo_queries	hotchpotch/miracl-hf-unified	119	921	119
zh	zh_queries	hotchpotch/miracl-hf-unified	200	1700	200

BM25 nDCG@10

nDCG@10 is computed from the included BM25 ranking against the included qrels.

Nano split	Tokenizer	Forced BM25 positives	BM25 nDCG@10
ar	stemmer:ar/arabic	19	0.5445
bn	whitespace:bn	9	0.5103
de	stemmer:de/german	39	0.3665
en	stemmer:en/english	6	0.5432
es	stemmer:es/spanish	16	0.5110
fa	whitespace:fa	9	0.5337
fi	stemmer:fi/finnish	14	0.6240
fr	stemmer:fr/french	55	0.3034
hi	stemmer:hi/hindi	6	0.5497
id	stemmer:id/indonesian	16	0.5705
ja	wordseg:ja	3	0.5956
ko	wordseg:ko	5	0.5090
ru	stemmer:ru/russian	31	0.4457
sw	whitespace:sw	9	0.5782
te	whitespace:te	8	0.6044
th	wordseg:th	7	0.6475
yo	whitespace:yo	13	0.5323
zh	wordseg:zh	18	0.4466

Skipped Tasks

None.

License

NanoMIRACL is a derived dataset. Users must comply with the licenses, terms, and attribution requirements of the upstream datasets listed above.