ViHiX: Vietnam High-school Examination Benchmark

A multimodal (text + image) benchmark of 3,006 questions from 84 official exam papers of Vietnam's national high-school examinations (the THPT Quốc gia exam in 2018–2019 and the Tốt nghiệp THPT graduation exam in 2020–2026), covering 13 subjects. The questions are in Vietnamese. Each question contains the text transcribed verbatim from the original paper, an answer checked against the official answer key published by the Ministry of Education and Training (MoET), cropped images for the 199 questions that rely on a figure, graph, diagram or table, and, for the 96 Geography questions that require the Vietnam Atlas, crops of the relevant Atlas map region (plus the matching symbol-legend rows when the page has no legend).

Intended uses:

• Evaluating (multimodal) language models on real Vietnamese exam questions, scored against official answers.
• Generating and filtering chain-of-thought data using the official answers as ground truth.
• Evaluating document and exam extraction from scanned PDFs.

Statistics

Questions per subject and year. For every (year, subject) pair we include one representative exam code (ma_de), since the different codes of the same exam are permutations of the same questions.

Subject	2018	2019	2020	2021	2022	2023	2024	2025	2026	Total
English	50	50	50	50	50	50	50	40	40	430
Mathematics	50	50	50	50	50	50	50	22	22	394
History	40	40	40	40	40	40	40	28	28	336
Biology	40	40	40	40	40	40	40	28	28	336
Chemistry	40	40	40	40	40	40	40	28	28	336
Physics	40	40	40	40	40	40	40	28	28	336
Geography	40	40	40	40	40	40	40	28	28	336
Civic Education (GDCD)	40	40	40	40	40	40	40	–	–	280
Informatics	–	–	–	–	–	–	–	30	30	60
Economics and Law Education (GDKTPL)	–	–	–	–	–	–	–	28	28	56
Literature (Ngữ văn)	–	6	6	6	6	6	6	7	7	50
Technology (Agriculture track)	–	–	–	–	–	–	–	28	–	28
Technology (Industry track)	–	–	–	–	–	–	–	28	–	28

• Question types: mcq: 2,820, true_false: 76, short_answer: 60, essay: 50
• Questions with images: 295 (199 with figures from the paper, 96 with Atlas crops: 135 map crops and 33 legend crops).
• Questions usable as ground truth (gt_usable): 2,909 (listed in cot_ready.jsonl).
• Subjects that rely on images the most: Technology (Industry track) ~43% of questions, Geography ~34% (5% figures, 29% Atlas crops), Mathematics ~17%, Biology ~14%, Physics ~11%, Chemistry ~5%. English, History, Civic Education, Economics and Law, Informatics, Technology (Agriculture track) and Literature are text-only.

Not included: foreign languages other than English (French, Russian, Chinese, German, Japanese, Korean).

Data format

Column	Description
id	<exam_id>_<part><number>, unique
exam_id, year, exam_type, subject, subject_name, ma_de	Exam, year, exam type (THPTQG for 2018–2019, TNTHPT for 2020–2026), subject (short code / Vietnamese name), exam code
global_no, part, q_no	Running index 1..N within the exam, section (I/II/III), and the question number printed on the paper
q_type	mcq (four options), true_false (four statements a–d), short_answer, essay (Literature)
passage	Shared reading passage or data (repeated for every question that uses it)
question	Question text (formulas as LaTeX inside $...$)
option_A..D, statement_a..d	Answer options (mcq) or statements (true_false)
answer	mcq: A–D; true_false: Đ,S,Đ,S (Đúng/true, Sai/false, in the order a–d); short_answer: a number string as in the official key (decimal comma; may be several values such as "8000; 8002"); essay: null
answer_status	official (MoET key) or none (not available or deliberately withheld)
key_debatable	true for 8 questions where the official key may be debatable: an LLM solving blind chose the same other answer in all 5 independent attempts, even though the transcription and the key were re-checked against the original paper. The official answer is kept as ground truth
gt_usable	true if the question can be used as ground truth: it has an official MoET answer and no known problem (excludes the 50 Literature essays, Geography 2021, 3 questions under review and 4 whose figure is missing from the source paper). 2,909 questions
image, images / image_paths	Parquet: embedded images (image = first image). JSON/CSV: paths inside images/. Images are figures from the paper and, for Geography questions that cite the Vietnam Atlas, crops of the relevant Atlas region (atlas_*_k.jpg) and symbol legend (atlas_*_L.jpg)
source_file, page, notes	Source file, page, and extraction notes (for example printing errors in the original paper)

Formats: data/*.parquet (images embedded, for load_dataset), data.jsonl, data.json and data.csv (images referenced through image_paths inside the images/ folder).

from datasets import load_dataset
ds = load_dataset("quin210/ViHiX", split="train")
ds = ds.filter(lambda x: x["gt_usable"])

Worked solutions

solutions.jsonl holds step-by-step worked solutions in Vietnamese for all 2,909 gt_usable questions. One line per question: {"id", "solution", "answer", "model_answer", "confidence", "atlas_free", "answer_given", "key_debatable", "doubt"?}.

Subset	Questions	How the solution was produced
answer_given = false	2,897	LLM solved the question without seeing the answer (for the 81 Atlas questions, reading the Atlas page); kept only because its final answer matches the official MoET answer. For hard questions up to 5 independent blind attempts were made
answer_given = true	12	The model got these wrong in all 5 blind attempts (the transcription and the official key were re-checked against the original paper). The solution was then written with the official answer given, explaining why it is correct and why the tempting options are wrong. model_answer keeps the model's earlier blind answer

• Median solution length ~710 characters; calculation subjects state the data, formulas and intermediate results, and theory subjects discuss every option or statement a–d.
• Answer matching normalizes the decimal comma/point; true_false must match all four statements.
• doubt (1 question, answer_given): the model found the official answer hard to justify (Biology 2023 Q113 depends on reading a graph value).
• key_debatable (8 questions, all answer_given): the model chose the same other answer in all 5 blind attempts, which may indicate a debatable official key; the official MoET answer is still used as ground truth.
• atlas_free is kept for compatibility and is false everywhere now that Atlas images are attached. The source Atlas scan is ~1170×1650 px per page, so very small labels are sometimes hard to read; such solutions say when they fall back on general knowledge.
• confidence is the model's own estimate (high / medium / low).
• For the cleanest training data, use answer_given == false (2,897 questions).

sol = load_dataset("quin210/ViHiX", data_files="solutions.jsonl", split="train")

Evaluation protocol

The eval/ folder contains the official scorer and protocol:

• eval/PROTOCOL.md: protocol v1.0 (input, prompts, decoding, metrics).
• eval/score.py: scorer. Run python eval/score.py predictions.jsonl --data cot_ready.jsonl with one {"id": ..., "answer": ...} per line.
• eval/prompts.json: fixed zero-shot prompts for each question type.

Metrics: score_pct (primary; ALL_excl_key_debatable is also reported) uses official-style partial credit, where true_false follows the MoET rubric (1 correct statement = 0.1, 2 = 0.25, 3 = 0.5, 4 = 1.0), mcq and short_answer are 0/1; exact_pct is strict exact match. A missing or unparsable answer counts as 0. Reports are broken down by subject, question type, exam type, image vs text-only, and year.

How the dataset was built and verified

1. The papers (PDF/DOC/DOCX, many of them low-quality scans) were converted to page images. A vision model transcribed each page verbatim, including printing errors in the original, and figures or tables were cropped by coordinates and checked one by one.
2. Answers were taken from the images of the official MoET answer keys, using the column of the right exam code. "Reference answers" published by third parties were not used as ground truth. When a key file was missing from the collection, the official table was found online (re-published by several news outlets) and cross-checked. An older OCR-derived answer set was found to contain many errors and was discarded.
3. Answer keys were read twice independently. Every key was read at least twice by separate passes (the second pass did not see the first pass's answers) and compared programmatically; every disagreement was resolved by re-reading the zoomed-in official table and checking the content of the question. This process corrected 5 wrong answers.
4. Independent spot check. A separate group of reviewers re-read the exam images before looking at the JSON for all doubtful questions plus 6 random questions per exam (about 480 questions). About 1% had small transcription errors (a wrong word, an extra crop), all fixed.
5. Solution-disagreement review. Every question whose model solution disagreed with the official answer was re-checked against the original paper. This found and fixed transcription errors (swapped options, missing "not", wrong numbers, a wrong figure) and a systematic problem in English: many questions had lost the printed instruction line (e.g. "CLOSEST/OPPOSITE in meaning") and underlines. All English questions (2018–2026) were then re-checked against the originals and fixed. No official answer was changed except one disputed key resolved from the MoET table image.
6. Every exam passes an automatic schema check (all questions present, all options present, valid answers, images exist, warnings when a question mentions a figure but has no image).

Source and license

The exam papers and answer keys are documents of Vietnam's Ministry of Education and Training, collected from public sources. The Atlas crops are short excerpts of the Atlat Địa lí Việt Nam (Vietnam Education Publishing House), included for education and research. The dataset (transcriptions, annotations, flags and worked solutions) is released under CC BY-NC 4.0: non-commercial use with attribution. Rights to the original exam papers and the Atlas remain with their owners.