This dataset contains extracted excerpts of Portuguese national exam content published by Instituto de Avaliação Educativa, I.P. (IAVE). We are not aware of an explicit redistribution license from IAVE for this content — see the "License & attribution" section below. Access is gated to keep this to research and educational NLP use; briefly tell us your intended use.

AMALIA IAVE national exams — verified MCQ pairs (2024-2025)

Verified multiple-choice question+answer pairs extracted from Portugal's 2024 and 2025 national secondary-school exams (Ensino Secundário, 12th grade), built for specializing AMALIA-9B-0626-DPO toward a K-12 tutor use case. Ground truth by construction: every answer is read directly off the official IAVE marking scheme (critérios de correção), never inferred by a model. Full pipeline, methodology, and the rest of the AMALIA improvement work: github.com/teex-pt/pt-amalia.

This is a complementary companion to teex-pt/amalia-pilot-honesty-v2 — that dataset is synthetic (templated + on-policy + teacher-drafted, all Apache-2.0 sources), this one is real exam content, which changes what you can assume about licensing (see below).

Why this exists

Two research groups have already built benchmarks from IAVE's national exam archive: pt_exams and PHEB (LREC 2026, from the same team behind AMALIA-9B), covering 2006–2023 across six subjects (Mathematics, Portuguese, History, Geography, Biology/Geology, Philosophy). This dataset is 2024–2025 only, across 26 subjects — disjoint from both by construction, so it can be used for training (e.g. LoRA fine-tuning, SFT) without contaminating evaluation on those benchmarks.

Scope

Source: iave.pt (Instituto de Avaliação Educativa, I.P.), Portugal's national exam authority. URLs verified live, not pattern-guessed.
Years: 2024, 2025. Sessions: 1ª Fase, 2ª Fase, Época Especial.
Subjects: 26, including Matemática A/B, Português, Filosofia, História A/B, Economia A, Física e Química A, Biologia e Geologia, and foreign languages (Alemão, Espanhol, Francês, Inglês, Italiano, Mandarim).
122 exam sittings parsed from 244 official PDFs (exam paper + marking scheme pairs).
270 verified MCQ question+answer pairs → 425 training records (two-version exams, which reshuffle option order per version, produce one record per version — 383 train / 42 valid). A further 37 items across 11 exam sittings are held out entirely (not published here) for this project's own internal harness evaluation — see "Held-out evaluation slice" below.

Not covered: open-response items (richer content, but graded by multi-level rubrics rather than a simple answer key — different, harder extraction problem, left for a follow-up). Ensino Básico (grades 1–9) exam streams are a separate archive not yet scraped.

Files

extracted.jsonl — one record per verified MCQ item: year, subject, code, phase, item, question (full stem + lettered options), answer_v1/answer_v2 (correct letter per exam version; answer_v2 is null for single-version exams), points, notation_risk (bool, see below).
mix/train.jsonl, mix/valid.jsonl — the same content reshaped into the project's standard chat schema ({"messages": [{"role": "user", ...}, {"role": "assistant", "content": "(X)"}]}), ready to drop into an SFT/LoRA run. 9:1 split, seeded independently from other AMALIA mixes.
extract-report.json — per-subject yield (items found in marking scheme vs. successfully paired with question text), for auditing extraction quality per subject rather than trusting one aggregate number.

Methodology

Download exam + marking-scheme PDF pairs from verified IAVE URLs (respectful rate limiting, resumable).
Parse marking schemes (three formats observed: answer tables, inline "N pontos" + lone letter, inline + "Versão 1/2" dual answers).
Locate each item's question text in the exam PDF and pair it with its answer. MCQ items only — an item is kept only if at least two lettered options ((A), (B), ...) appear in its matched text, a safety gate against a real bug this exercise caught (see below).
Convert to the training schema.

99% pairing yield (270/272 MCQ items found in marking schemes were successfully matched to their question text).

Two real bugs found and fixed during extraction (disclosed, not swept under the rug)

Item-numbering collisions. Exams are organized into groups (GRUPO I/II/III, PARTE A/B/C) whose item numbers restart per group, so "item 4" can legitimately refer to different items in different groups. An early version of the extractor matched only the first occurrence of an item number and, in one caught case, mislabeled an open-response Português item with an unrelated MCQ answer. Fixed by requiring ≥2 lettered options in the matched span (a cheap, high-precision safety check) and scanning all occurrences of the item number, keeping the first one that passes.
Page-footer bleed-through. pdftotext -layout linearizes PDF pages back-to-back, so page footers ("Prova 501/1.ª F. • Página 12/ 15") and the next page's header text were leaking into the tail of captured questions in 90/265 records (34%) before this was caught by manual spot-checking and fixed with a targeted strip.
Stray control characters. 60/270 records (22%) carried a \x07 (BEL) character from a PDF bullet-glyph that didn't map to a printable character. Cosmetic, not a ground-truth error, but common enough to fix rather than leave as noise — verified safe to strip unconditionally.

Held-out evaluation slice

Before splitting into mix/train.jsonl/mix/valid.jsonl, the build script reserves 11 whole exam sittings (37 items, 11 distinct subjects: Alemão, Biologia e Geologia, Espanhol, História A, Italiano, Latim A, Mandarim, Matemática A, Matemática Aplicada, Português Língua Segunda) for this project's internal harness evaluation — not included in this dataset's files. This is a real fix, not a stylistic choice: an earlier version of the split was row-level only, and empirically leaked — every exam sitting represented in valid also had other items from the same sitting in train, including 37 cases where a question's two shuffled-option versions (same stem, different correct letter) ended up split across both sides. mix/valid.jsonl here is fine for its actual purpose (SFT trainer loss monitoring), but reusing it as a held-out benchmark would have been memorization dressed up as generalization. The reservation is deterministic (smallest sittings first, capped at one sitting per subject code, so no single large sitting like Economia A's 56-item slate can dominate the holdout or gut that subject's training presence).

Known limitations

Math notation risk. pdftotext's font handling can mangle calculus/ algebra notation (fractions, limits, derivatives) in STEM items — verified by manual inspection, not an automatic detector. Records from Matemática A/B, Matemática Aplicada, Física e Química A, and Geometria Descritiva A carry notation_risk: true (102/425, 24%) so they can be filtered or reviewed separately. Many items in these same subjects (combinatorics, geometry-by-figure, plain-number problems) extract cleanly — the flag is a coarse per-subject signal, not a per-item guarantee.
MCQ only. Open-response items are excluded from this release entirely.
Secondary education only. No Ensino Básico (grades 1–9) coverage yet.

License & attribution

The extraction code and dataset structure (this repo's pipeline, schema, and this card) are shared under the same terms as the rest of teex-pt/pt-amalia (Apache-2.0).

The question text and answers themselves are drawn from IAVE's official national exam papers, published by Instituto de Avaliação Educativa, I.P., a Portuguese public institute, at iave.pt for public access. We are not aware of an explicit redistribution license from IAVE for this content. We publish this extracted, structured excerpt on the same basis as PHEB — the AMALIA consortium's own peer-reviewed (LREC 2026) benchmark, which redistributes the same category of content (raw question text from these same IAVE archives, 2006–2023) as a public research artifact. If you have a specific commercial or redistribution use case for this data, we'd recommend confirming directly with IAVE; this release is intended for research and educational NLP use, consistent with that precedent.

Citation

@misc{amalia-iave-2026,
  title = {AMALIA IAVE national exams: verified MCQ pairs (2024-2025)},
  author = {teex-pt},
  year = {2026},
  howpublished = {\url{https://huggingface.co/datasets/teex-pt/amalia-iave-exams-2024-2025}},
  note = {Question content sourced from Instituto de Avaliação Educativa, I.P. (IAVE), Portugal}
}

Related work this builds on: PHEB (Tavares et al., LREC 2026).

Downloads last month: 8

Total file size:

562 kB