Madras1/rag-qa-fulltext-ptbr

RAG QA Full-Text PT-BR Mistral

A large-scale dataset of Brazilian Portuguese RAG-style question-answer pairs with grounded evidence spans, generated from Madras1/corpus-ptbr-v1 documents using Mistral models. Every answer is anchored to literal quotations from the source text, making this dataset suitable for training and evaluating retrieval-augmented generation systems, extractive QA models, and reading comprehension benchmarks in Portuguese.

Two configurations are available:

• qa_pairs (recommended) — one row per QA pair; each row is self-contained with the full source context, the question, the answer, and the evidence_spans that support it.
• documents — one row per source document with nested QA pairs, useful when you want to work at the document level.

How to Use

Load the QA pairs (default)

from datasets import load_dataset

ds = load_dataset("Madras1/rag-qa-fulltext-ptbr", "qa_pairs", split="train")
print(ds[0]["question"])
print(ds[0]["answer"])
print(ds[0]["evidence_spans"])

Load the documents

ds_docs = load_dataset("Madras1/rag-qa-fulltext-ptbr", "documents", split="train")
doc = ds_docs[0]
print(doc["context"][:200])
for qa in doc["qa_pairs"]:
    print(f"Q: {qa['question']}")
    print(f"A: {qa['answer']}")

Stream without downloading

ds = load_dataset(
    "Madras1/rag-qa-fulltext-ptbr",
    "qa_pairs",
    split="train",
    streaming=True,
)
for row in ds:
    print(row["question"], "→", row["answer"])
    break

Dataset Summary

Metric	Value
Raw JSONL rows	462,414
Deduplicated documents	434,394
QA pairs	993,423
Documents with QA	374,871
Documents without QA	59,523
Removed duplicate rows	28,020
Evidence spans capped at	3

Domain Distribution

Domain	Documents		Domain	Documents
política	45,652		ciência	15,043
jurídico	42,483		negócios	14,876
saúde	39,628		governo	14,298
economia	38,538		segurança	11,963
tecnologia	36,756		entretenimento	10,314
educação	21,441		esportes	10,296
religião	20,280		história	9,315
meio ambiente	20,040		agro	9,009
geral	19,835		outros (12)	34,627
cultura	15,998

Document Type Distribution

Type	Documents		Type	Documents
notícia	197,389		opinião	6,790
artigo	76,367		regulamento	5,313
artigo educativo	40,303		manual	4,854
outro	33,691		tutorial	4,107
blog	21,780		decreto	2,755
verbete	12,417		lei	2,120
paper	10,465		FAQ	1,916
guia	7,284		relatório	6,843

Models Used

Model	Documents	Share
mistral-small-latest	310,857	71.6%
mistral-medium-latest	108,448	25.0%
mistral-small-2603	11,920	2.7%
mistral-large-latest	2,696	0.6%
mistral-large-2512	473	0.1%

Example

A single row from the qa_pairs config looks like this:

{
  "qa_id": "doc-12345-qa-0",
  "document_id": "doc-12345",
  "hf_row_idx": 12345,
  "qa_index": 0,
  "question": "Qual é o principal objetivo da Lei de Acesso à Informação?",
  "answer": "Garantir o acesso a informações públicas por qualquer pessoa, sem necessidade de justificativa.",
  "context": "A Lei nº 12.527, de 18 de novembro de 2011, conhecida como Lei de Acesso à Informação (LAI), regulamenta o direito constitucional de acesso dos cidadãos às informações públicas...",
  "evidence_spans": [
    "regulamenta o direito constitucional de acesso dos cidadãos às informações públicas"
  ],
  "evidence_text": "regulamenta o direito constitucional de acesso dos cidadãos às informações públicas",
  "evidence_span_count_original": 1,
  "evidence_truncated": false,
  "source": "Madras1/corpus-ptbr-v1",
  "subset": "real",
  "domain": "juridico",
  "doc_type": "artigo",
  "model": "mistral-small-latest"
}

Note: This is an illustrative example. Actual field values vary per row.

Schemas

qa_pairs

Field	Type	Description
qa_id	string	Stable QA identifier (doc-{idx}-qa-{n}).
document_id	string	Stable source document identifier.
hf_row_idx	int64	Original row index from Madras1/corpus-ptbr-v1.
qa_index	int64	QA pair index within the parent document.
question	string	Generated question in Brazilian Portuguese.
answer	string	Generated answer grounded in the source context.
context	string	Full source document text.
evidence_spans	list[string]	Literal quotations copied from context that support the answer.
evidence_text	string	Evidence spans joined by double newline for convenience.
evidence_span_count_original	int64	Evidence count before capping at 3.
evidence_truncated	bool	Whether the evidence list was capped.
source	string	Source dataset identifier.
subset	string	Source subset name.
word_count	int64	Word count of the source document.
char_count	int64	Character count of the source document.
language	string	Language tag of the source document.
domain	string	Classified domain (e.g. juridico, saude, tecnologia).
doc_type	string	Classified document type (e.g. noticia, artigo, lei).
provider	string	API provider used for generation.
model	string	Model name used for generation.
generated_at	string	ISO 8601 timestamp of when the QA pair was generated.

documents

Field	Type	Description
document_id	string	Stable source document identifier.
hf_row_idx	int64	Original row index from Madras1/corpus-ptbr-v1.
context	string	Full source document text.
source	string	Source dataset identifier.
subset	string	Source subset name.
word_count	int64	Word count of the source document.
char_count	int64	Character count of the source document.
language	string	Language tag of the source document.
domain	string	Classified domain.
doc_type	string	Classified document type.
qa_pairs	list[struct]	Nested list of QA pairs, each with question, answer, evidence_spans, evidence_span_count_original, and evidence_truncated.
qa_count	int64	Number of QA pairs for this document.
provider	string	API provider used for generation.
model	string	Model name used for generation.
generated_at	string	ISO 8601 timestamp of generation.

Provenance

• Source corpus: Madras1/corpus-ptbr-v1
• Source subset: real
• Source field: c4_pt
• Annotation provider: Mistral API (mistral-small-latest, mistral-medium-latest, mistral-large-latest)
• Annotation method: Structured JSON output with response_format.json_schema

The source corpus is distributed as ODC-By 1.0 on Hugging Face. This derived dataset preserves attribution fields and should be used with the same care around source data limitations, web-crawl noise, and generated annotation quality.

Data Processing

1. Each source document was sent to the Mistral API with a prompt requesting domain classification, document type classification, and up to 3 QA pairs with literal evidence spans.
2. Evidence spans were validated as exact substrings of the source text during generation. Spans that were not literal matches were discarded.
3. Questions with fewer than 8 characters, answers with fewer than 8 characters, or QA pairs with no valid evidence were filtered out.
4. When the same source document appeared multiple times (28,020 duplicates), the best version was selected by preferring: more QA pairs → stronger model → most recent generation timestamp.
5. Evidence spans were capped at 3 per QA pair in the final dataset. The original count is preserved in evidence_span_count_original.

Limitations

• Synthetic annotations: QA pairs are generated automatically by Mistral models and may contain factual errors, hallucinated content, or low-quality questions.
• Evidence grounding: Evidence spans are verified as literal substrings of the source text during generation, but factual correctness depends on the source text itself.
• Empty documents: 59,523 source documents (13.7%) produced no usable QA pairs and are present only in the documents config.
• Deduplication: Duplicate raw rows were deduplicated by hf_row_idx. When duplicates existed, the row with the most QA pairs from the strongest model was kept.
• Model mix: The majority of annotations (71.6%) come from mistral-small-latest. Quality may vary across the different Mistral model tiers used. The model field in each row allows filtering by model.