ViEduHop: Vietnamese Educational Multihop Reasoning Benchmark with Textbook Provenance

ViEduHop is an evaluation-only benchmark of 3,015 Vietnamese questions on the grade 10-12 school textbooks Kết nối tri thức với cuộc sống (Connecting Knowledge with Life). Every question that has an answer carries gold source spans: the book, lesson, printed page and span id of the passage(s) the answer needs, so a system can be scored on where it found its evidence as well as on what it answered. It accompanies the paper "OTTER: Verifiable Multihop Reasoning over Educational Ontology for Vietnamese Education" (ACIIDS 2027 submission).

Subcategories and tiers

The 3,015 rows form four subcategories (one config each) made of nine tiers (tier field). Scores of different subcategories or tiers must not be averaged into one number.

subcategory (config)	tiers (rows)	format	what it measures
S1 Textbook QA (s1_textbook_qa, 1,646)	core 1,164; cross 191; extra 291	4-option multiple choice, 1-3 hops, one subject or several	understanding and applying 1-3 textbook passages
S2 Multi-hop Reasoning (s2_multihop, 175)	multihop (3 hops 66, 4 hops 44, 5 hops 65)	4-option multiple choice	gathering 3-5 passages from several lessons or subjects, mostly calculation or date reasoning
S3 Abstention (s3_abstention, 330)	abstain 300; twist_abstain 30	4 options plus option E "Không đủ căn cứ trong sách" (the book does not provide enough evidence) added at evaluation time	recognising that the book does not support an answer; half of the rows are answerable controls
S4 Robustness and Formats (s4_robustness, 864)	twist 36; para 228; typed 600	multiple choice; true/false (375); fill-in (150); MCQ (75)	rewording, designed traps and other question formats

Formats: mcq4 2,160, mcq4_abstain 330, true_false 375, fill_blank 150. Subjects: Physics 349, Chemistry 555, Biology 467, History 600, Geography 522, Literature 522 rows (a cross-subject row counts under its first subject). Grades: 10: 826, 11: 798, 12: 1,391 rows. Excluding the 228 Para rows, which paraphrase existing items, there remain 2,787 rows (2,718 distinct question-option pairs; an unanswerable Abstain item and its answerable control share one stem).

One example per subcategory (English glosses in brackets)

• S1 (VIEDUHOP-C2-HIS-017, History 12, 2 hops): Khoảng cách thời gian giữa giai đoạn kháng chiến toàn dân chống Pháp (1951 – 1953) và chiến thắng Ấp Bắc (1963) là bao nhiêu năm? [How many years separate the nationwide resistance against the French, 1951-1953, from the Ap Bac victory, 1963?] A. 10 B. 5 C. 8 D. 15 years; key A.
• S2 (VIEDUHOP-M4-0021, 4 hops, Chemistry 10 and 11, Geography 10, Physics 12): a person of 70 kg is 3% nitrogen by mass; the same atoms as N₂ gas at standard conditions fit in which volume of air (78.1% N₂)? A. 4.30 B. 1.68 C. 8.03 D. 2.15 m³; key D.
• S3 (OTTER-ABS-BIO-MPR-001, Biology 11, missing premise): Khi một người bị suy tim đi từ tư thế đứng sang tư thế nằm xuống, điều gì có thể xảy ra với huyết áp và vận tốc máu? [When a heart-failure patient goes from standing to lying down, what may happen to blood pressure and blood velocity?] Options A-D are pairs of changes; the right answer is E, because the book does not treat heart failure.
• S4: Twist VIEDUHOP-TW-BRG-PHY-001 (a quantity defined through the mass and volume of a gas can be negative; in which scale? key: Celsius); true/false TF_048_3 (Không đúng khi cho rằng nguyên lí chuyển dịch cân bằng phụ thuộc vào thay đổi áp suất và số mol khí. [It is not correct that Le Chatelier's principle depends on pressure and moles of gas.] key: False); fill-in FIB_059 (Trong điện phân, sự _____ xảy ra tại cathode. [In electrolysis, _____ occurs at the cathode.] key: khử (reduction)).

How to load

from datasets import load_dataset
ds = load_dataset("quin210/vieduhop", "all", split="test")
headline = ds.filter(lambda r: r["headline"])                      # 1,208 items of the pre-registered headline

The Hub shows one split, test, per config: the whole benchmark is meant for evaluation. The original split of each row is in original_split (dev, test, extra, multihop, twist, eval_only). Do not train or tune on these rows; if you tune, use rows with original_split = "dev" and report on the rest.

Data fields

field	type	meaning
id	string	unique item id
tier, subcategory, format	string	tier (core, cross, extra, multihop, abstain, twist_abstain, twist, para, typed), config name, answer format
original_split, headline, hard_tier, seen_in_development	string, bool	original split; headline = the 1,208 Core/Cross test items that were never used to develop the reference system; hard_tier = 136-item post hoc subset of the headline set; seen_in_development = used while developing the reference system (excluded from the headline)
subject (Vietnamese), subject_en, subjects_en, grade, topic	string, list, int	subject(s) of the item; topic only for Typed
question	string	the question (Vietnamese)
choices, choice_labels	list	options and their letters; empty for fill-in; true/false uses "Đúng" / "Sai"
answer	string or null	key letter; the answer text for fill-in; null for unanswerable S3 rows (the key is the added option E)
answerable	bool	false only for unanswerable S3 rows
hop_count, hop_type, difficulty, bloom, book_specific	int, string, bool	number of gold spans the answer needs; reasoning type (single, bridge, comparison, temporal, calculation, causal_chain, conditional_rule ...); difficulty and Bloom level (where available; not for every tier); book_specific = flagged by the closed-book probe and evidence checks (392 of the 477 flagged rows are answered by neither reference model above half of four option permutations without the book; the other 85 were kept on the evidence checks alone)
gold_span_ids, book_ids, lesson_keys	list	ids of the gold spans (<book>:p<page>:u<k>), books and lessons
source_pointers	list of struct	for each gold span: span_id, book_id, book_title, grade, lesson (heading path), printed_page, pdf_page; absent for Abstain and Typed rows (ids only)
twist_type, abstain_type, control_type, twin_id, orig_id, alignment_status, negated, closed_book_hard, extra_kind	mixed	tier-specific metadata: Twist type; Abstain type (missing premise, out of corpus, counterfactual guard, almost enough) and its answerable control and twin; the original item of a paraphrase; Typed span alignment (auto_aligned, key_conflict, not_aligned); Multi/Extra closed-book flag

Not included, on purpose: textbook passages, gold quotes, explanations (they may quote the book), generator and verifier outputs.

Evaluation protocol

scripts/evaluate.py (no dependencies) scores a prediction file. The pre-registered headline metrics on the 1,208 headline items are accuracy (abstention counts as wrong), located (the top returned source span is a gold span), lesson_located, verified-correct (correct and located) and cost. Everything else (Hard, Twist, Para, Typed, Extra, Multi, per-subcategory tables, abstention F1) is additional and was fixed after the main results were known. For S3 give the system the extra option E and score abstention precision, recall and F1 on the unanswerable rows. The exact definitions are in the paper.

Ethical considerations and copyright

The textbooks belong to their publisher and are not redistributed; this dataset contains question text, options, keys, metadata and pointers (book, lesson heading, page, span id). The pointers refer to the authors' internal corpus; users need a legal copy of the textbooks to read the passages. No personal data is included (teachers are not named in the data). The benchmark is for research on retrieval, provenance and abstention; it must not be used to grade or rank students or schools.

License

The questions, options, keys, metadata and source pointers of ViEduHop are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International licence (CC BY-NC-SA 4.0; https://creativecommons.org/licenses/by-nc-sa/4.0/). The choice follows the terms of the source material: the textbook files used to write the questions are sample editions that their publisher offers free of charge for teaching and learning, not for commercial use. The textbooks and their text are not part of this dataset and remain the property of their publisher; the evaluation scripts are under the Apache License 2.0. This is the authors' choice, not legal advice.

Files

data/*.jsonl (the five configs), docs/DATASET_CONSTRUCTION.md, LICENSE, CHECKSUMS.sha256, stats.json, scripts/ (evaluate.py, load_example.py, validate_release.py, upload_to_hub.py).

Version history

• v2 (frozen 28 Oct 2026): 3,015 rows; source hash of the unified file in stats.json.