README.md

---
pretty_name: rag-eval-ptbr — Brazilian tax-law RAG evaluation golden set
license: cc-by-4.0
language:
  - pt
task_categories:
  - question-answering
  - text-retrieval
tags:
  - rag
  - rag-evaluation
  - llm-as-judge
  - legal
  - tax
  - brazil
  - portuguese
size_categories:
  - n<1K
configs:
  - config_name: tributario-golden
    data_files:
      - split: test
        path: tributario-golden/test.parquet
  - config_name: tributario-corpus
    data_files:
      - split: test
        path: tributario-corpus/test.parquet
---

# rag-eval-ptbr — golden set tributário (pt-BR)

End-to-end **RAG evaluation** golden set in **Brazilian Portuguese**, tax-law domain
(CTN — National Tax Code — and LC 123/2006, Simples Nacional). Designed to measure
faithfulness/groundedness, context precision/recall, retrieval quality and citation
accuracy of RAG pipelines — not a reading-comprehension QA set.

Built for (and consumed by) the open-source harness
[`rag-eval-ptbr`](https://github.com/rafaelsouza-tech/rag-eval-ptbr), which adds a
statistically grounded regression runner (`rageval run` / `rageval compare`).

*Versão em português abaixo.*

## Configs

| Config | Rows | What it is |
|---|---|---|
| `tributario-golden` | 35 | evaluation items (split `test`) |
| `tributario-corpus` | 536 | canonical article-level corpus with stable semantic IDs |

**Golden item fields:** `id`, `question` (natural question, as a taxpayer/lawyer would ask),
`reference_answer` (curated, fully supported by the labeled passages),
`golden_passage_ids` (complete and minimal set of supporting corpus chunks),
`reference_citations` (canonical format `(CTN, art. 173, I)`), `answerable`
(4 items are deliberately unanswerable from the corpus — abstention probes) and
`metadata` (question type: factual / multi_hop / condicional / temporal / nao_respondivel;
difficulty; source law; smoke-set flag; curation provenance).

**Corpus chunk fields:** `doc_id` (semantic, hierarchical, stable — `ctn:art-173`,
`lc123:art-18-p4`), `text` (verbatim official text), `source_law`, `article`, `url`.
Scope: full CTN (arts. 1–218) + LC 123/2006 (arts. 1–41). Retrieval metrics
(recall@k, MRR, nDCG) are deterministic `doc_id` matching — no LLM needed.

## Construction & curation

1. Corpus parsed verbatim from official Planalto sources by a reproducible script
   (`scripts/build_corpus.py` in the repo); articles over ~350 tokens subdivided by paragraph.
2. Items **LLM-generated** grounded on real corpus passages, by a model from a different
   family than the judge (against self-preference bias).
3. **Adversarial verification, one independent pass per item**: support of every answer
   claim checked against the actual chunk text, completeness/minimality of
   `golden_passage_ids`, citation format, question naturalness. Items were fixed or dropped.
4. Deduplication and consolidation (36 → 35 items).

**Honest provenance:** `curated_by` reads `geração sintética por LLM (verificação
adversarial; revisão humana pendente)`. Item-by-item human review is the acceptance bar for v1.0;
this v0.x release is published for reproducibility and early use.

## Limitations

- Single domain (federal tax law); 35 items — designed for paired regression testing,
  not leaderboard-scale comparison.
- Reference answers reflect the **statute text as compiled in July 2026**; they ignore
  case law by design (answers must be fully supported by the corpus).
- Generated by an LLM: despite adversarial verification, residual errors are possible —
  open an issue in the repo if you find one.
- Not legal advice.

## License & sources

Source texts are official Brazilian normative acts, **in the public domain**
(Law 9.610/1998, art. 8, IV). The curation/annotation layer (questions, reference
answers, labels) is released under **CC-BY-4.0**. Attribution: *Rafael Souza,
rag-eval-ptbr golden set tributário*.

## Citation

```bibtex
@misc{souza2026rageval,
  author = {Souza, Rafael},
  title = {rag-eval-ptbr: RAG evaluation harness and golden set for Brazilian Portuguese},
  year = {2026},
  url = {https://huggingface.co/datasets/rafaelsouza-tech/rag-eval-ptbr}
}
```

---

# Versão em português

Golden set de **avaliação de RAG ponta a ponta** em **português brasileiro**, domínio
tributário (CTN e LC 123/2006). Feito para medir faithfulness/groundedness, context
precision/recall, qualidade de retrieval e acurácia de citação de pipelines RAG — não é
um QA set de compreensão de leitura.

Construção: corpus verbatim das fontes oficiais do Planalto (script reprodutível);
itens gerados por um LLM ancorados em passagens reais; **verificação adversarial item a
item**; deduplicação (36 → 35). Proveniência honesta: revisão
humana item a item ainda pendente (barra de aceitação da v1.0).

4 itens são deliberadamente **não-respondíveis** pelo corpus (sondas de abstenção).
As citações seguem o formato canônico `(CTN, art. 173, I)`. Os `doc_id` do corpus são
semânticos e estáveis (`ctn:art-173`) — métricas de retrieval são casamento
determinístico de IDs, sem LLM.

Textos-fonte em domínio público (Lei 9.610/98, art. 8º, IV); camada de curadoria em
**CC-BY-4.0**. Limitações: domínio único; 35 itens (desenhado para teste de regressão
pareado); reflete a legislação compilada em julho/2026; ignora jurisprudência por
construção; não é aconselhamento jurídico.