NanoMTEB-BR

This dataset is a Nano-style Brazilian Portuguese retrieval benchmark for HAKARI-Bench. It is derived from the retrieval and reranking tasks exposed by MTEB-BR at commit e17f36ae304e7bdb76202744c5fb56b440050213. The six public splits are Quati, JurisTCU, BRTaxQAR, FaQuADIR, MedPTRetrieval, and FaqBacenRetrieval.

Usage

from datasets import load_dataset

dataset_id = "hakari-bench/NanoMTEB-BR"
split = "MedPTRetrieval"

queries = load_dataset(dataset_id, "queries", split=split)
corpus = load_dataset(dataset_id, "corpus", split=split)
qrels = load_dataset(dataset_id, "qrels", split=split)
bm25_candidates = load_dataset(dataset_id, "bm25", split=split)
dense_candidates = load_dataset(dataset_id, "harrier_oss_v1_270m", split=split)
reranking_candidates = load_dataset(dataset_id, "reranking_hybrid", split=split)

Data Layout

This dataset uses the same six Hugging Face Datasets configs as the current HAKARI-Bench Nano sets:

• corpus: documents with _id and text
• queries: queries with _id and text (the default config)
• 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 dense candidates

Every config exposes the same six split names. Qrels are positive-only; source rows with score > 0 are collapsed to binary positive qrels. Rows with non-positive scores are excluded from qrels and may be used as hard-negative corpus candidates. These binary Nano qrels are therefore not a graded copy of every full MTEB-BR judgment.

Candidate Construction

• bm25: local BM25 top-500, capped by corpus size, with the resolved tokenizer recorded per split.
• harrier_oss_v1_270m: dense top-500, capped by corpus size, from microsoft/harrier-oss-v1-270m at revision 31de22b673913c7d658c0f03f792d77c2dcf8ebd. Queries use the web_search_query prompt, documents use no prompt, embeddings are normalized, and scoring is FP32 cosine similarity.
• reranking_hybrid: reciprocal-rank fusion over BM25 and dense rankings with rrf_k=100, retaining the RRF top-100.

If the RRF top-100 contains no qrels-positive document, one deterministically selected positive is appended at rank 101. The safeguard is outside nDCG@10 and Recall@100.

Nano Construction and Sampling

The Nano build selects at most 200 eligible queries and at most 10,000 corpus documents per split, while always retaining the selected positive documents. Quati and JurisTCU keep every eligible query and use their pinned reranking candidate sources as hard-negative provenance. Their hard-negative documents are sampled with deterministic query round-robin before source-order corpus fill. BRTaxQAR and FaqBacenRetrieval use seed-42 stable-hash sampling. FaQuADIR maximizes positive-paragraph and title coverage. MedPTRetrieval uses seed-42 stratification across its seven question types.

Query and document aliases are deduplicated within each split using canonical comparison keys built with NFKC normalization, whitespace collapse, and case folding. These comparison keys do not rewrite the stored source text. The first stable-sampling query is the query representative. The first source-order document by default is the document representative. BRTaxQAR has narrowly audited, pin-scoped preferred canonical corpus IDs for exact duplicate-body groups whose first source labels are demonstrably incorrect. Qrels that reference removed aliases are remapped to those representatives; when a positive and non-positive judgment collapse to the same query-document pair, the positive judgment wins.

BRTaxQAR and FaQuADIR compare document aliases by body text so different source titles do not preserve duplicate answer bodies as unjudged alternatives. The stored representative title and body remain source-native; this rule only controls equivalence and identifier remapping.

The pinned MTEB-BR/faquad-ir wrapper omits five contexts from the original FaQuAD data and maps the 14 questions attached to those contexts to unrelated documents. FaQuADIR cross-checks the wrapper against liafacom/faquad at commit 6ad978f20672bb41625b3b71fbe4a88b893d0a86, restores the five missing contexts verbatim, and remaps only those 14 qrels before alias deduplication and q200 sampling. Repair decisions and the original file SHA-256 are recorded in metadata/FaQuADIR.json.

For MedPTRetrieval, this release retains the complete 500-document MTEB-BR corpus and does not add documents mined from the 384,095-row parent AKCIT/MedPT dataset. The final 200 queries are sampled across the seven question-type strata; those types are sampling metadata, not a classification task. MTEB-BR contains repeated question text with different source answers, so all known answers within the 500-pair MTEB-BR sample that are attached to an identical canonical source question are unioned into qrels. This produces 204 positive qrels for 200 queries.

Quati and JurisTCU use documents from their pinned reranking candidate datasets as hard-negative corpus provenance. The associated reranking tasks are folded into these retrieval splits rather than exposed as additional public splits. NanoMTEB-BR is not registered in the built-in benchmark collections and does not alter the standard 551-task evaluation set.

JurisTCU reads the pinned raw ENUNCIADO values and preserves their HTML-like tags after trimming only surrounding whitespace. This intentionally differs from the pinned MTEB-BR loader's tag-stripping transformation so that redistributed source markup remains benchmark input. Tags are ignored only when building document-alias comparison keys; the stored representative text is not rewritten.

Split Statistics

Character statistics use 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
Quati	49	10000	892	48.6	47.0	55.0	1108.6	1181.0	1257.0
JurisTCU	150	10000	1711	61.4	50.0	93.0	314.6	285.0	382.0
BRTaxQAR	200	461	546	131.4	120.5	161.2	30656.5	4511.0	26996.0
FaQuADIR	200	243	201	62.3	57.0	71.0	825.6	772.0	1021.0
MedPTRetrieval	200	500	204	152.0	112.5	200.2	368.2	278.0	451.2
FaqBacenRetrieval	200	1654	201	70.5	66.0	88.2	518.8	400.0	651.0

Candidate Quality

nDCG@10 and macro Recall@100 are computed against the included positive qrels and reported on a 0-100 scale. The optional rank-101 safeguard is excluded from both cutoffs.

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	-	51.78	60.90	58.52	82.12	87.11	92.00	-	24
Quati	regex	38.15	64.94	52.02	66.61	95.17	91.01	100	0
JurisTCU	regex	46.88	52.27	54.90	69.57	72.61	78.49	100	0
BRTaxQAR	regex	46.50	28.42	38.10	90.23	61.64	91.75	100-101	6
FaQuADIR	regex	88.70	84.25	87.79	100.00	99.50	100.00	100	0
MedPTRetrieval	regex	43.28	69.74	59.75	78.08	97.50	96.50	100-101	7
FaqBacenRetrieval	regex	47.17	65.77	58.58	88.25	96.25	94.25	100-101	11

Hybrid Safeguard Summary

• Safeguard positives: 24
• Rows limited by corpus size: 0
• Safeguarded query coverage: 100.00% for every split
• Safeguarded relevant coverage: Quati 88.90%, JurisTCU 77.91%, BRTaxQAR 91.03%, FaQuADIR 100.00%, MedPTRetrieval 100.00%, FaqBacenRetrieval 99.50%
• The safeguard guarantees one positive for a zero-hit query; it does not guarantee that every positive qrel is retained for multi-positive queries.
• Detailed generation metadata: reranking_hybrid_metadata.json

Benchmark and Dataset Sources

The benchmark task definitions are pinned to the MTEB-BR task suite. Dataset revisions and task-specific papers or references are pinned separately so the materialized Nano data can be traced back to the original benchmark sources.

Nano split	MTEB-BR task	Primary dataset	Revision	Related hard-negative source	Paper/reference	Recorded upstream license
Quati	Quati	MTEB-BR/quati-50k	5cb87d9561d8ace807305f50e59a4b0af352da2e	MTEB-BR/quati-reranking	Oliveira et al. (2024) — Quati	cc-by-4.0
JurisTCU	JurisTCU	LeandroRibeiro/JurisTCU	ac7bea9e580626a586ffea69b245d26f5a73d44e	MTEB-BR/juristcu-reranking	Fernandes et al. (2026) — JurisTCU	cc-by-4.0 as recorded by MTEB-BR from dataset-author confirmation; the pinned upstream card still contains an XXX placeholder
BRTaxQAR	BRTaxQAR	unicamp-dl/BR-TaxQA-R	9f0f4263928c6152a40f888da3e63375a24bb0f3	none	Domingos Júnior et al. (2025) — BR-TaxQA-R	cc-by-4.0
FaQuADIR	FaQuADIR	MTEB-BR/faquad-ir	51fd9e7707bb4971229a0189560379992c3adce2	none	Sayama et al. (2019) — FaQuAD	cc-by-4.0
MedPTRetrieval	MedPTRetrieval	AKCIT/MedPT	6a55a33e289539ad676da89dd20208729acc7509	none	Färber et al. (2025) — MedPT	cc-by-4.0
FaqBacenRetrieval	FaqBacenRetrieval	MTEB-BR/faq-bacen	076d89a68a8b8d2f14e3161631c416ffe29b8463	none	Finardi et al. (2023) — FAQ-BACEN	apache-2.0

Original Dataset and License Links

The wrapper-backed materialized sources above are supplemented with direct links to their original dataset or license authority:

• Quati: unicamp-dl/quati original dataset
• FaQuADIR: liafacom/faquad original dataset, FaQuAD CC BY 4.0 license
• FaqBacenRetrieval: Itau-Unibanco/FAQ_BACEN original dataset

Known Limitations and Publication Notes

• Retrieval qrels are inherited annotations, not a new expert relabeling. Source datasets can contain weak or non-answering source-positive pairs, and unjudged documents are not proven non-relevant. NanoMTEB-BR preserves those judgments unless a pinned, reproducible source-mapping defect is established.
• Quati and JurisTCU source judgments are graded. NanoMTEB-BR intentionally collapses every source judgment with score > 0 to one binary positive. Consequently, low-grade topical matches and direct answers are evaluated as equally positive; this is not a preservation of the original graded scale.
• Returning MedPTRetrieval to the MTEB-BR 500-document corpus removes the additional false-negative risk introduced by the retired parent-corpus augmentation. It does not create exhaustive semantic relevance judgments: a medically plausible answer attached to another source question can remain unjudged. Such alternatives are source-inherited false negatives under the positive-only qrels.
• BRTaxQAR retains each non-empty source document's full pinned filedata after trimming surrounding whitespace. No 32,000-character construction cap is applied, so qrels-relevant passages late in long legal documents remain present in the published corpus. The Harrier dense candidate encoder has a 32,768-token maximum sequence length, and evaluated encoders have their own limits. Full corpus retention therefore prevents corpus-materialization-time character truncation but does not guarantee that a late passage influences every document embedding. It also cannot restore content when the pinned source file itself is incomplete: the Constituição Federal de 1988 snapshot ends after Article 24. Of its references, 20 affected source qrel pairs require later provisions such as Article 40 or Article 150; five occur in the selected q200. The corpus snapshot remains source-native, but those qrel pairs are excluded by exact pin-scoped decisions because the evaluated document text does not contain their cited provisions.
• BRTaxQAR validates recurring Solução de Consulta, Ato Declaratório, and Parecer Normativo citations because their document numbers can be reused or collide across instrument families. Reference titles and source-document headers are compared by instrument family, subtype, number, and year. Confident mismatches are omitted from qrels. Day- or month-only differences and unparseable headers are retained with audit warnings rather than silently removed.
• BRTaxQAR also applies exact pin-scoped audit decisions for source mappings that are unsafe to infer with a broad parser, including named-instrument collisions and one ambiguous embedded-reference expansion. Irrelevant targets are omitted. A verified source-label typo that points to the correct, byte-identical legal body is instead remapped and retained, preventing a new false negative. The preferred canonical corpus IDs for duplicate-body aliases with misleading labels are selected against the primary body header. In this materialization, 69 reference occurrences were rejected, corresponding to 66 query-document qrel pairs. 12 retained references produced audit warnings, and 1 verified source-label typo was retained after exact alias correction. The pin-scoped manifest matched 35 reference occurrences. 3 audited duplicate-body groups use preferred canonical corpus IDs. Full decisions, evidence, group membership, and body hashes are recorded in metadata/BRTaxQAR.json. Source queries 356 and 614 lost every qrel under this validation and were excluded before q200 sampling.
• Source-native markup, control characters, replacement characters, and spacing artifacts can occur in the pinned source text. They are intentionally retained as benchmark input rather than silently cleaned or rewritten.
• Quati has 49 queries because one of the 50 source queries has no positive judgment after the score > 0 policy.

Structural Audit Status

The packaged files passed the automated structural and candidate-integrity checks recorded in audit.json. These checks cover schemas, identifiers, qrel references, candidate uniqueness, and candidate coverage. They do not certify that inherited relevance judgments are semantically exhaustive or that source text is free of source-native noise.

License

NanoMTEB-BR follows the same mixed licensing model as MTEB-BR. It is a mixed-license benchmark compilation and is not distributed under one blanket license:

• Redistributed corpus documents, queries, and qrels retain each original dataset's license and attribution requirements. NanoMTEB-BR does not relicense them; consult the source table and pinned links above.
• NanoMTEB-BR-authored documentation, sampling metadata, and candidate rankings follow MTEB-BR's results-dataset license, CC BY 4.0.
• The HAKARI-Bench software used to build and evaluate this dataset is under the MIT License; MTEB-BR's own benchmark code is Apache-2.0.

JurisTCU is recorded as CC BY 4.0 by the pinned MTEB-BR task metadata, based on confirmation from dataset author L. Fernandes on 2026-06-08. The pinned original JurisTCU card still contains an unreplaced XXX license placeholder. This release relies on MTEB-BR's documented author confirmation and does not independently relicense JurisTCU.

Citation

When using NanoMTEB-BR, cite MTEB-BR and the original source for every task used. The pinned MTEB-BR README provides this citation:

@misc{mteb-br-2026,
  title  = {MTEB-BR: A Text Embedding Benchmark for Brazilian Portuguese},
  author = {Stekel, Tardelli R. C.},
  year   = {2026},
  doi    = {10.5281/zenodo.21087217},
  url    = {https://doi.org/10.5281/zenodo.21087217}
}

Machine-readable Provenance

• metadata/<split>.json: source revisions, sampling, corpus construction, and completed candidate-generation metadata
• audit.json: structural and candidate integrity audit
• reranking_hybrid_metadata.json: BM25, Harrier, RRF, safeguard, and quality metadata