This dataset is derived from official Portuguese legislation text (public, copyright-exempt — see License below). Access is gated only so we can track who's using it and for what — briefly tell us your intended use.

AMALIA cita-legal — grounded legal citation SFT (RAG-first)

(question + real source excerpts → answer that cites the exact article, or a refusal when the excerpts don't answer the question). Built for specializing AMALIA-9B toward Portuguese legal text, as the grounded-answering counterpart to teex-pt/amalia-sum-dre. Derived from teex-pt/leis-pt-consolidada by teex-pt/pt-amalia.

Why this exists (RAG-first, not closed-book)

leis-pt's own project spec concludes that in a legal domain, where hallucinating is more costly than in an exam, it makes more sense to train "a model that's good at using a search tool" than "a model that memorized the law" — and this project's own IAVE LoRA pilots back that up empirically: closed-book fact-recall SFT on small legal/exam data moved answer format, not precision (rejected at 0pp on the mcq target; a second pass reached only +5.4pp, still not a clean pass). So this dataset does not train closed-book legal QA. Instead it trains the same answer-only-from-context-or-refuse contract leis-pt's own retrieval service already enforces in production, deliberately kept consistent with it rather than reinvented.

Ground truth by construction

Questions are real, official summaries of amending laws (e.g. "Procede à segunda alteração à Portaria n.º 46/2015...") describing what each law actually changed — not synthesized by a model.
Answers are templated, extractive citations of the real target article's text — no free-text generation, no model call anywhere in the pipeline, so there's no hallucination risk to inherit.
Refusals pair a real question with excerpts from a genuinely unrelated diploma, answered with a fixed refusal string consistent with leis-pt's own production refusal behavior.

Scope

7,538 grounded (positive) examples, each with up to 6 source excerpts.
1,130 refusal (negative) examples (~15% of the positive count) — real question, deliberately wrong excerpts.
8,222 train / 446 valid, split diploma-disjoint (no diploma's examples appear on both sides) and balanced so a handful of heavily-amended codes (Código Civil, Código Penal, ...) can't dominate one split — applying the lesson from this project's own IAVE sitting-level leak (see that dataset's card) proactively rather than repeating it.
A small fraction of source amendment records (~4%) couldn't be resolved to a specific article and were dropped rather than guessed — see build-report.json.

Files

train.jsonl, valid.jsonl — {"messages": [{"role": "user", "content": "PERGUNTA: ...\n\nEXCERTOS:\n[F1] ...«...»\n..."}, {"role": "assistant", "content": "<grounded citation answer, or the refusal string>"}]}. Each record also carries target_diploma_id, label (grounded/refusal), n_fragments_available, n_fragments_used.
build-report.json — event/resolution/split yield stats.

Known limitations

A small fraction of citations may point to the wrong instance of an article when a diploma has more than one fragment sharing the same article label (rare) — folded into the same ~4% noise budget above, not separately quantified.
Answers are templated citations, not natural-language explanations. This trains citation discipline and grounding, not conversational legal explanation — deliberately, to keep every answer verifiable against its source with zero generation risk.
valid.jsonl is for SFT trainer loss monitoring, not a retrieval or QA benchmark — leis-pt's own hand-validated benchmark work ("Stage B") is tracked separately in that project, not here.
Decontaminated against all four consortium benchmarks (2026-07-14, 13-word shingle overlap check, 9,614 rows checked): LegalBenchPT and pt_exams flagged 53 + 2 rows respectively, all manually verified as benign shared real-document text (statutes and one shared government policy title), not leaked test items — no fictional exam content from either appears anywhere in this dataset. alba (linguistics MCQs) and cultura_viva (culture/trivia MCQs) came back fully clean, zero overlap. Full reports: eval/results/DECONTAMINATION-{legalbenchpt,pt_exams,alba-culturaviva}.md in teex-pt/pt-amalia.

License & attribution

Same basis as teex-pt/leis-pt-consolidada: official Portuguese legal texts are copyright-exempt (CDADC art.º 8.º). This compilation is released under CC0 1.0.

Citation

@misc{amalia-cita-legal-2026,
  title = {AMALIA cita-legal: grounded Portuguese legal citation SFT pairs},
  author = {teex-pt},
  year = {2026},
  howpublished = {\url{https://huggingface.co/datasets/teex-pt/amalia-cita-legal}},
  note = {Questions and target text sourced from Diário da República's official records; templated, not model-generated}
}

Downloads last month: 26

Total file size:

50.6 MB