Part of the VixSTORY: Vietnamese University Admission QA collection. Released for non-commercial research under CC BY-NC-SA 4.0. Access is granted automatically once you submit this form.

By requesting access, you agree to use VixSTORY for non-commercial research only, not to redistribute it outside the terms of its license, and to cite the paper 'A Benchmark for Structured Multihop Reasoning in Cross-Institution University Admission Advisory Question Answering' (AMI at PAKDD 2026) in any work that uses it. Citing the paper is a condition of access.

VixSTORY: Vietnamese Cross-Institution University Admission Advisory QA

VixSTORY (Vietnamese Cross-InStitution UniversiTy AdmissiOn AdvisoRY QA) is a benchmark for structured multihop reasoning over real university admission regulations in Vietnamese. Answering a question requires combining evidence from 1 to 5 regulatory contexts, either within one institution or across several institutions whose policies differ in terminology, structure and style.

The benchmark targets retrieval-augmented generation (RAG) systems, from lexical and dense retrievers to iterative and graph-based RAG. It ships with a closed retrieval corpus, so it can be used both for end-to-end QA and for evidence retrieval.

Paper: Long S. T. Nguyen, Tin T. Ngo, Dung N. H. Le, Quynh T. N. Vo, Dung H. Nguyen, Khang N. Le, and Tho T. Quan. A Benchmark for Structured Multihop Reasoning in Cross-Institution University Admission Advisory Question Answering. Workshop on Autonomous Machine Intelligence: From Theory to Practice (AMI), at PAKDD, 2026. [Website]

TL;DR: VixSTORY contains 4,499 Vietnamese multihop QA pairs (1–5 hops) on university admission regulations within and across six universities in Ho Chi Minh City, together with a corpus of 3,462 regulatory passages. In the paper, graph-based RAG methods such as HippoRAG 2 clearly outperform flat retrieval and remain more robust as the number of hops increases.

Part of a collection: VixSTORY is featured in the VixSTORY: Vietnamese University Admission QA collection on Hugging Face.

Citation required: Access is granted on the condition that any work using this dataset cites the paper. See the BibTeX entry below.

1. Dataset at a glance

VixSTORY pairs multihop admission questions with a closed corpus of regulations from six universities:

Property	Value
QA pairs	4,499 (train 2,553 · validation 246 · test 1,700)
Retrieval corpus	3,462 regulatory passages
Institutions	6 universities in Ho Chi Minh City
Hop levels	1–5
Reasoning scope	within-institution, cross-institution
Question types	comparative, factual, procedural, verification, conditional, bridge
Language	Vietnamese

2. Usage

VixSTORY is a gated dataset. Before loading it:

Open this page while signed in to Hugging Face and submit the access form. Access is granted immediately.
Authenticate locally with hf auth login, or pass token=... to load_dataset.

Once access is granted, both configurations can be loaded with the datasets library:

from datasets import load_dataset

qa = load_dataset("ura-hcmut/VixSTORY", "qa")              # train / validation / test
corpus = load_dataset("ura-hcmut/VixSTORY", "corpus", split="corpus")

sample = qa["test"][0]
print(sample["question"])
print(sample["answer"])
print(sample["context_ids"])      # gold supporting passages, resolvable in `corpus`

Evaluation settings: VixSTORY can be used in two ways.

Open-book RAG (main setting in the paper): index the corpus config, retrieve for each test question, and generate an answer. Retrieval quality is measured against context_ids (Recall@k).
Reading comprehension / oracle setting: give the model the gold contexts directly, which isolates multihop reasoning from retrieval.

The test split is the evaluation set used in the paper, so results should be reported on it; train and validation are intended for development, few-shot exemplars or fine-tuning.

3. Data fields

VixSTORY is released as two configurations: qa holds the question–answer pairs, and corpus holds the passages they are grounded in.

3.1 `qa` config

Each question–answer pair comes with its reasoning metadata and gold evidence:

Field	Type	Description
`id`	string	Unique instance identifier
`question`	string	Advisory question in Vietnamese
`answer`	string	Reference free-form answer grounded in the gold contexts
`hop_level`	int	Number of reasoning hops (1–5)
`question_type`	string	`comparative`, `factual`, `procedural`, `verification`, `conditional` or `bridge`
`scope`	string	`within` (one institution) or `cross` (two or more institutions)
`topic`	string	Primary admission topic (see below)
`school_ids`	list[string]	Institutions the question is about
`school_names`	list[string]	Full Vietnamese names of these institutions
`context_ids`	list[string]	Gold supporting passage IDs (all present in the `corpus` config)
`contexts`	list[struct]	Gold supporting passages inline: `context_id`, `school_id`, `title`, `text`

3.2 `corpus` config

Each passage is stored together with its source institution and document:

Field	Type	Description
`context_id`	string	Passage ID, formatted `CTX-<SCHOOL>-<DOC_HASH>-CH<chunk>`
`school_id`	string	Institution code
`school_name`	string	Full Vietnamese institution name
`title`	string	Title or slug of the source document
`text`	string	Passage text

Passages that share CTX-<SCHOOL>-<DOC_HASH> come from the same source document.

3.3 Example (test split, contexts truncated)

The following two-hop question compares the regulations of HCMUT and UEH:

{
  "id": "q164",
  "question": "Trường Đại học Bách khoa và Đại học Kinh tế TP. Hồ Chí Minh có quy định thời gian cho các hoạt động học tập và thi cử khác nhau như thế nào?",
  "answer": "Trường Đại học Bách khoa quy định một tiết học là 50 phút. Đối với Đại học Kinh tế TP. Hồ Chí Minh, tổng thời gian trình bày và trả lời câu hỏi của giám khảo cho mỗi thí sinh trong thi vấn đáp không quá 15 phút.",
  "hop_level": 2,
  "question_type": "comparative",
  "scope": "cross",
  "topic": "chuong_trinh_dao_tao",
  "school_ids": ["HCMUT", "UEH"],
  "school_names": ["Trường Đại học Bách khoa, ĐHQG-HCM", "Trường Đại học Kinh tế TPHCM"],
  "context_ids": ["CTX-HCMUT-89E8D0E044-CH0003", "CTX-UEH-3D1F614E56-CH0025"],
  "contexts": [
    {"context_id": "CTX-HCMUT-89E8D0E044-CH0003", "school_id": "HCMUT", "title": "Quy_dinh_dao_tao_Thac_si", "text": "... Để hoàn thành khối lượng kiến thức 1 tín chỉ, học viên phải cần thêm 2 – 4 tiết chuẩn bị, tự học mỗi tuần (ngoài giờ lên lớp). Một tiết học được quy định bằng 50 phút. ..."},
    {"context_id": "CTX-UEH-3D1F614E56-CH0025", "school_id": "UEH", "title": "QD4419_CongTacKhaoThi", "text": "Tổng thời gian trình bày và trả lời câu hỏi của giám khảo không quá 15 phút cho mỗi thí sinh. Điểm thi vấn đáp phải được công bố công khai ngay sau mỗi buổi thi ..."}
  ]
}

4. Splits and statistics

Overview: VixSTORY contains 4,499 QA pairs in total.

Split	Train	Validation	Test	All
Instances	2,553	246	1,700	4,499

By hop level: most training questions need one or two hops, while the test split concentrates on three-hop questions.

Hops	Train	Validation	Test	All
1	1,126	117	212	1,455
2	799	70	206	1,075
3	269	26	883	1,178
4	204	19	196	419
5	155	14	203	372

By question type: comparative and factual questions are the most common, followed by procedural, verification and conditional ones.

Type	Train	Validation	Test	All
Comparative	853	60	921	1,834
Factual	908	100	167	1,175
Procedural	258	27	307	592
Verification	374	47	47	468
Conditional	148	12	250	410
Bridge	12	0	8	20

By reasoning scope: more than half of the test questions require evidence from several institutions.

Scope	Train	Validation	Test	All
Within-institution	1,759	190	724	2,673
Cross-institution	794	56	976	1,826

The test split emphasizes cross-institution and higher-hop reasoning, with longer questions and answers on average (172 vs. 103 characters per question and 401 vs. 253 per answer, compared with train).

Institutions: the corpus covers six universities, with HCMUT and UEL contributing the most passages.

Code	Institution	Passages
HCMUT	Ho Chi Minh City University of Technology, VNU-HCM (Trường Đại học Bách khoa)	956
UEL	University of Economics and Law, VNU-HCM (Trường Đại học Kinh tế - Luật)	941
HCMUS	University of Science, VNU-HCM (Trường Đại học Khoa học Tự nhiên)	510
HCMUSSH	University of Social Sciences and Humanities, VNU-HCM (Trường Đại học Khoa học Xã hội và Nhân văn)	485
UEH	University of Economics Ho Chi Minh City (Trường Đại học Kinh tế TP.HCM)	433
FTU	Foreign Trade University, Ho Chi Minh City Campus (Trường Đại học Ngoại thương)	137

Topics: academic programs are by far the most frequent topic, followed by admission methods and tuition.

Topic	Meaning	Instances
`chuong_trinh_dao_tao`	Academic programs	1,859
`phuong_thuc_xet_tuyen`	Admission methods	560
`hoc_phi`	Tuition	433
`quy_che_quy_dinh`	Academic regulations	331
`nghe_nghiep_viec_lam`	Careers and employment	207
`hoc_bong_ho_tro`	Scholarships and financial aid	202
`thu_tuc_ho_so`	Application procedures	193
`ngoai_ngu`	Foreign-language requirements	128
`diem_chuan`	Cut-off scores	119
`thong_tin_chung`	General information	82
`hoc_phi_hoc_bong`	Tuition and scholarships	54
`tin_tuc_su_kien`	News and events	50
`sau_dai_hoc`	Graduate studies	33
`khac`	Other	248

Question types: each question belongs to one of six types.

Comparative: contrasts a policy across programs or institutions.
Factual: asks for a specific fact stated in the regulations.
Procedural: asks how to complete an admission or academic process.
Verification: checks whether a claim or eligibility condition holds.
Conditional: asks what applies under a given condition.
Bridge: needs an intermediate entity to link two pieces of evidence.

5. Construction

VixSTORY is built with a controlled semi-automatic pipeline:

Guideline design: Annotators and a domain expert define admissible question types, evidence-grounding requirements and multihop constraints.
Corpus collection: Admission-related regulations are crawled from the official websites of the six universities. They are then normalized, de-noised and chunked into atomic regulatory passages with institution provenance.
Intra-institution semantic clustering: Passages are embedded and clustered per institution to approximate latent admission topics.
Cross-institution multihop context selection: Topic clusters are aligned across institutions. Tuples of h contexts are sampled within or across institutions under semantic relatedness constraints.
QA generation and validation: An LLM generates question–answer pairs from each tuple with an intermediate reasoning trace. Outputs are validated for multihop completeness, evidence coverage and logical consistency, and ambiguous cases go to expert review.

Data splits: the released splits were created as follows.

The test split is the evaluation set reported in the paper.
Train and validation are drawn from the remaining instances (90/10, seed 42), stratified by hop level and reasoning scope, with no question shared across splits.
All gold context_ids resolve to passages in the corpus config.

6. Benchmark results

The table below reports the main results from the paper on the test split. All systems use GPT-4o-mini for generation and text-embedding-3-small for embeddings. Judge is an LLM-as-a-Judge score from GPT-4o.

Method	F1	Recall@5	Recall@10	Judge
BM25	0.268	0.204	0.248	0.213
Dense	0.318	0.240	0.294	0.235
Hybrid	0.360	0.298	0.315	0.296
IRCoT	0.366	0.355	0.441	0.342
MiniRAG	0.476	0.419	0.487	0.400
RAPTOR	0.509	0.479	0.512	0.444
LightRAG	0.639	0.537	0.564	0.603
HippoRAG 2	0.656	0.614	0.697	0.644

Graph-based methods clearly outperform flat and reasoning-guided retrieval, and their advantage is most pronounced on high-hop and cross-institution questions, highlighting the value of explicit structure for regulation-driven advisory QA.

7. Intended use and limitations

VixSTORY is designed for research, so please keep the following points in mind:

Intended use: research on multihop QA, RAG, graph-based retrieval and advisory assistants for Vietnamese. VixSTORY is a benchmark, not an authoritative source of admission information.
Temporal validity: regulations, tuition and cut-off scores change every admission cycle. Answers reflect the documents at collection time (2024–2026 admission cycles) and may be outdated. Always consult official university channels.
Question style: questions and answers are LLM-generated from real regulations and validated for faithfulness, so their phrasing may differ from that of real applicant queries.
Coverage: the benchmark covers six major universities in Ho Chi Minh City.

8. License

Released under CC BY-NC-SA 4.0. The source texts are public regulations published by the respective universities, and their rights remain with those institutions. The dataset is for non-commercial research only.

9. Citation

If you use VixSTORY, please cite:

@inproceedings{10.1007/978-981-92-2014-4_19,
  title        = {{A Benchmark for Structured Multihop Reasoning in Cross-Institution University Admission Advisory Question Answering}},
  author       = {Nguyen, Long S. T. and Ngo, Tin T. and Le, Dung N. H. and Vo, Quynh T. N. and Nguyen, Dung H. and Le, Khang N. and Quan, Tho T.},
  year         = 2027,
  booktitle    = {Trends and Applications in Knowledge Discovery and Data Mining},
  publisher    = {Springer Nature Singapore},
  address      = {Singapore},
  pages        = {235--247},
  isbn         = {978-981-92-2014-4},
  editor       = {Long, Cheng and Shi, Jieming and Liang, Yuxuan and Huang, Xiao and Choi, Byron and Song, Yuanfeng and Yiu, Man Lung},
}

10. Contact

For questions, issues or feedback, please contact Long S. T. Nguyen (long.nguyencse2023@hcmut.edu.vn), URA Research Group, Faculty of Computer Science and Engineering, Ho Chi Minh City University of Technology (HCMUT), VNU-HCM.

Downloads last month: 44

Total file size:

9.51 MB

Request access to VixSTORY