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 sum-dre — Portuguese legal act summarization (SFT)

(diploma full text → official sumário) instruction-tuning pairs, built for specializing AMALIA-9B toward Portuguese legal text. Ground truth by construction: every target summary is Diário da República's own official abstract (sumário) for the act, never inferred by a model. Derived from teex-pt/leis-pt-consolidada by teex-pt/pt-amalia.

Why this exists

This project's own IAVE LoRA pilots (see JOURNAL.md in pt-amalia) showed closed-book fact-recall SFT on small legal/exam data has a low ceiling — format moved, precision didn't. Summarization is a different kind of task: it trains compression and extraction given the source text right there in context, not recall of facts the model has to have memorized. It's also the task PLANO.md's leis-pt spec names first (amalia-sum-dre) as "gold by construction."

Scope

1,006 diplomas, capped to those whose full consolidated text fits a single training context window (≤32,000 characters). The 721 diplomas above that cap — mostly the 26 base legal codes and a handful of omnibus reform laws — are a different task (retrieval source, not a single-summary target); see teex-pt/leis-pt-consolidada and teex-pt/amalia-cita-legal instead.
956 train / 50 valid, split by diploma (each diploma contributes exactly one row, so this is diploma-disjoint by construction — no leakage to guard against).

Files

train.jsonl, valid.jsonl — the project's standard chat schema: {"messages": [{"role": "user", "content": "<instruction + diploma header + full text>"}, {"role": "assistant", "content": "<official sumário>"}]}. Each record also carries diploma_id, tipo, numero, designacao, fonte_url, n_chars_input for traceability/filtering.
build-report.json — yield stats (kept vs. skipped-too-long vs. skipped-empty).

Known limitations

valid.jsonl is for SFT trainer loss monitoring, not a benchmark — it's a random diploma-level split, not curated for difficulty or topic balance.
The official sumário is itself sometimes terse (a single sentence) — this trains extractive/compressive summarization in the DR abstract's own house style, not a general-purpose summarizer.
Decontaminated against all four consortium benchmarks (2026-07-14, 13-word shingle overlap check): LegalBenchPT and pt_exams flagged rows all manually verified as benign shared real-document text (statutes, and one shared government policy title), not leaked test items; alba and cultura_viva 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 (including their official DR abstract) are copyright-exempt (CDADC art.º 8.º). This compilation is released under CC0 1.0.

Citation

@misc{amalia-sum-dre-2026,
  title = {AMALIA sum-dre: Portuguese legal act summarization SFT pairs},
  author = {teex-pt},
  year = {2026},
  howpublished = {\url{https://huggingface.co/datasets/teex-pt/amalia-sum-dre}},
  note = {Target summaries sourced from Diário da República's official sumário field}
}

Downloads last month: 24

Total file size:

16.6 MB