Datasets:

igor-eduardo-research
/

mirage-pt

Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
query_id: string
query: string
specialty: string
complexity: string
country: string
language: string
doc_id: string
source_type: string
title: string
publisher: string
to
{'doc_id': Value('string'), 'title': Value('string'), 'source_type': Value('string'), 'language': Value('string'), 'country': Value('string'), 'publisher': Value('string')}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
                  return get_rows(
                      dataset=dataset,
                  ...<4 lines>...
                      column_names=column_names,
                  )
                File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
                  return func(*args, **kwargs)
                File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
                  rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
                File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
                  yield from ds.decode(False) if ds.features else ds
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2951, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2461, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2486, in _iter_arrow
                  for key, pa_table in self.ex_iterable._iter_arrow():
                                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 547, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 430, in _iter_arrow
                  for key, pa_table in self.generate_tables_fn(**gen_kwags):
                                       ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              query_id: string
              query: string
              specialty: string
              complexity: string
              country: string
              language: string
              doc_id: string
              source_type: string
              title: string
              publisher: string
              to
              {'doc_id': Value('string'), 'title': Value('string'), 'source_type': Value('string'), 'language': Value('string'), 'country': Value('string'), 'publisher': Value('string')}
              because column names don't match

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

MIRAGE-PT: Portuguese Clinical RAG Benchmark

A public benchmark for evaluating retrieval-augmented generation systems on Portuguese clinical text.

Overview

MIRAGE-PT provides 500 synthetic clinical queries in Brazilian Portuguese across 6 medical specialties. Query–document relevance labels were generated by two LLM judges (inter-judge Cohen's κ = 0.954; the judge models are not named in v1.0). No human or physician relevance judgments are included in v1.0. A two-physician validation of a stratified 120-pair sample is in progress, and results will be reported here. The corpus references 25 publicly available Brazilian clinical documents (Ministério da Saúde, CONITEC, ANVISA/CMED, DATASUS, RENAME). No patient data is included.

Stat	Value
Queries	500
Specialties	6
Relevance judgments	1,487 lines in `qrels.txt` (1,065 unique query–document pairs; see Known issues)
Corpus documents	25
Language	Brazilian Portuguese
Annotation	LLM-generated (2 LLM judges, inter-judge κ=0.954); physician validation in progress

Specialties

Specialty	Simple	Complex	Total
Pharmacology	97	—	97
Clinical Decision	40	60	100
Emergency	57	23	80
Pricing/Coverage	75	—	75
Pediatrics	48	22	70
Clinical Protocols	45	25	70

Files

File	Description
`queries.jsonl`	500 synthetic clinical queries with specialty and complexity labels
`corpus.jsonl`	Metadata only (ID, title, source type, publisher) for 25 public Brazilian clinical references; document text not included
`qrels.txt`	TREC-format LLM-generated relevance labels (1,487 lines; 1,065 unique query–document pairs)
`evaluate.py`	Evaluation script (Recall@k, P@5, NDCG@5)

Quick Start

from datasets import load_dataset

ds = load_dataset("igor-eduardo-research/mirage-pt")

Evaluation

python3 evaluate.py --results your_results.jsonl

Results format (your_results.jsonl):

{"query_id": "Q0001", "doc_ids": ["DOC0003", "DOC0007", "DOC0001"]}

Baselines

System	Recall@5	P@5	NDCG@5
BM25-only	99.0%	—	—
Dense-only (t=0.20)	77.8%	—	—
Dense-only (t=0.60)	20.2%	—	—
Hybrid+RRF (no authority)	95.6%	—	0.354
Hybrid+RRF+Authority	96.6%	29.8%	0.354

Submit your system's results by opening an issue with your results.jsonl.

Corpus Sources

All documents are publicly available Brazilian clinical references:

Ministério da Saúde (MS) — Clinical protocols, Cadernos de Atenção Básica
CONITEC — PCDTs (Protocolos Clínicos e Diretrizes Terapêuticas)
ANVISA / CMED — Regulatory and pricing references
DATASUS — CID-10 coding reference
RENAME 2024 — National essential medicines list

Known issues (v1.0)

qrels.txt has 1,487 lines but 1,065 unique query–document pairs (exact duplicate lines). After deduplication, 534 pairs are labeled relevant, which does not match the per-judge counts in the preprint (460 / 488). A corrected, documented qrels release is planned.
17 of 500 queries have no relevance labels.
corpus.jsonl contains metadata only (no document text), and several titles are file names. A source map with official URLs will be added.

Limitations

Queries are synthetic and may not reflect real physician information needs. Relevance labels are LLM-generated; agreement between two LLM judges measures consistency, not clinical validity. A physician validation is in progress.

Citation

@misc{eduardo2026miragept,
  title     = {BM25 and Dense Retrieval Are Complementary for Portuguese Clinical Text:
               An Empirical Study of Hybrid RAG Across 500 Clinical Queries},
  author    = {Eduardo, Igor},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.19686739},
  url       = {https://doi.org/10.5281/zenodo.19686739}
}

License

CC BY 4.0 — Igor Eduardo, 2026.

Downloads last month: 127

Total file size:

119 kB