Part of the ViHERMES: Vietnamese Healthcare Regulation 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 this dataset for non-commercial research only, not to redistribute it outside the terms of its license, and to cite the paper 'ViHERMES: A Graph-Grounded Multihop Question Answering Benchmark and System for Vietnamese Healthcare Regulations' (ACIIDS 2026) in any work that uses it. Citing the paper is a condition of access.

ViHERMES: Vietnamese Healthcare Regulations Multihop Reasoning Dataset

ViHERMES (Vietnamese HEalthcare Regulations – Multihop REasoning DataSet) is a benchmark for multihop question answering over Vietnamese healthcare regulations. Answering a question requires combining 1 to 5 legal units (articles or clauses) from different regulatory documents, following dependency patterns such as amendment tracing, cross-document comparison and procedural synthesis.

Every question comes with its supporting legal units, extractive evidence spans and an explicit reasoning explanation, and the release includes the retrieval corpus of regulatory contexts, so ViHERMES can be used for end-to-end RAG as well as for evidence retrieval.

Paper: Long S. T. Nguyen*, Quan M. Bui*, Tin T. Ngo, Quynh T. N. Vo, Dung N. H. Le, and Tho T. Quan. ViHERMES: A Graph-Grounded Multihop Question Answering Benchmark and System for Vietnamese Healthcare Regulations. Asian Conference on Intelligent Information and Database Systems (ACIIDS), 2026. [Website] [Code]

TL;DR: ViHERMES contains 1,561 Vietnamese multihop QA pairs (1–5 hops) on healthcare regulations, grounded in 998 legal units from 490 Ministry of Health documents. In the paper, a graph-aware system that models legal relations between units outperforms strong baselines, including HippoRAG 2 (F1 of 0.83 vs. 0.80).

Part of a collection: ViHERMES is featured in the ViHERMES: Vietnamese Healthcare Regulation 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

ViHERMES combines multihop questions with a corpus of healthcare regulations, summarized below:

Property	Value
QA pairs	1,561 (`test` split used in the paper)
Hop levels	1–5, balanced across levels
Retrieval corpus	998 legal units from 490 regulatory documents
Annotations	Gold context IDs, evidence spans and reasoning for every question
Domain	Vietnamese healthcare regulations
Language	Vietnamese

2. Usage

This 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/ViHERMES", "qa", split="test")
corpus = load_dataset("ura-hcmut/ViHERMES", "corpus", split="corpus")

sample = qa[0]
print(sample["question"])
print(sample["context_ids"])   # gold legal units, resolvable in the corpus

Evaluation settings: ViHERMES can be used in two ways.

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

3. Data fields

ViHERMES is released as two configurations: qa holds the questions with their annotations, and corpus holds the legal units they are grounded in.

3.1 `qa` configuration

Each question carries its gold legal units, evidence spans and reasoning:

Field	Type	Description
`id`	string	Unique question identifier
`question`	string	Multihop question in Vietnamese
`answer`	string	Reference answer grounded in all gold legal units
`hop_level`	int	Number of reasoning hops (1–5), equal to the number of gold legal units
`context_ids`	list[string]	Gold legal unit IDs, all present in the `corpus` configuration
`evidence`	list[string]	Extractive evidence span from each gold legal unit, in the same order as `context_ids`
`reasoning`	string	Explanation of how the evidence is combined into the answer
`context_block`	string	Gold legal units rendered as text (ID, title and body), ready to use as model input

3.2 `corpus` configuration

Each legal unit is an article or clause, stored with the title of its source document:

Field	Type	Description
`context_id`	string	Legal unit ID
`title`	string	Title of the regulatory document
`text`	string	Text of the legal unit (article or clause)

3.3 Example (evidence and context block truncated)

The following two-hop question links a 2024 decision on methadone treatment with a 2025 dispatch on Esketamine:

{
  "id": "vihermes_00252",
  "question": "Trong khoảng thời gian từ năm 2024 đến đầu năm 2025, Bộ Y tế đã ban hành những văn bản nào liên quan đến việc điều trị nghiện chất dạng thuốc phiện và quản lý một hoạt chất có nguy cơ bị lạm dụng?",
  "answer": "Trong giai đoạn 2024 đến đầu năm 2025, Bộ Y tế đã có các chỉ đạo quan trọng liên quan đến hai vấn đề này. Cụ thể, vào năm 2024, Bộ Y tế đã ban hành Quyết định về \"Hướng dẫn điều trị nghiện các chất dạng thuốc phiện bằng thuốc methadone\". Sau đó, vào ngày 20 tháng 01 năm 2025, Bộ Y tế (Cục Quản lý Dược) đã ban hành Công văn số 366/BYT-QLD về việc quản lý thuốc và nguyên liệu làm thuốc chứa hoạt chất Esketamine.",
  "hop_level": 2,
  "context_ids": [
    "3357-8894-4785",
    "1800-3811-8289"
  ],
  "evidence": [
    "Cung cấp thông tin về việc Bộ Y tế ban hành Quyết định về \"Hướng dẫn điều trị nghiện các chất dạng thuốc phiện bằng thuốc methadone\" vào năm 2024.",
    "Cung cấp thông tin về việc Bộ Y tế (Cục Quản lý Dược) ban hành Công văn số 366/BYT-QLD ngày 20/01/2025 về việc quản lý thuốc và nguyên liệu chứa hoạt chất Esket ..."
  ],
  "reasoning": "Để trả lời câu hỏi, người đọc cần tổng hợp thông tin từ cả hai văn bản. Văn bản 1 xác định hoạt động của Bộ Y tế trong năm 2024 liên quan đến điều trị nghiện bằng methadone. Văn bản 2 xác định hoạt động của Bộ Y tế vào đầu năm 2025 liên quan đến quản lý Esketamine. Câu trả lời đầy đủ yêu cầu kết hợp cả hai sự kiện này để tạo ra một cái nhìn tổng quan về các hoạt động của Bộ Y tế trong khoảng thời gian được nêu.",
  "context_block": "[CONTEXT 1]\\nID: ...\\nTitle: ...\\nBody: ..."
}

4. Statistics

Lengths are measured in characters and reported as min / mean / max.

Hops	Questions	Question length	Answer length	Context length
1	312	48 / 113.2 / 215	58 / 181.8 / 504	195 / 1,185.1 / 4,257
2	312	120 / 249.4 / 446	89 / 399.1 / 810	734 / 2,729.4 / 10,075
3	312	168 / 325.7 / 688	135 / 524.0 / 1,100	743 / 3,739.8 / 14,767
4	312	179 / 386.4 / 813	175 / 748.0 / 1,696	1,461 / 5,655.7 / 16,556
5	313	194 / 445.9 / 977	101 / 889.9 / 1,938	1,115 / 6,451.6 / 18,878
All	1,561	48 / 304.2 / 977	58 / 548.8 / 1,938	195 / 3,953.9 / 18,878

As the number of hops grows, questions and answers become longer and the gold evidence spans more and longer regulatory texts.

5. Construction

ViHERMES is built with a controlled pipeline that combines human supervision, semantic clustering and LLM-based generation:

Guideline design: annotators and a domain expert define multihop criteria, evidence usage and legal validity constraints.
Corpus collection and preprocessing: regulatory documents are crawled from official sources such as the public portals of the Vietnamese Ministry of Health, then normalized, cleaned and segmented by legal hierarchy into legal units, each stored as an (ID, title, text) triple.
Semantic clustering: legal units are embedded, normalized and clustered with spherical K-Means, and the units closest to each centroid are selected as representative contexts.
Multihop QA generation and validation: tuples of h contexts from different documents are sampled with bounded overlap and passed to an LLM that writes question–answer pairs relying on all h contexts, together with evidence spans and reasoning. Outputs are checked for format, hop–evidence consistency and explicit reasoning, and borderline cases are reviewed by annotators under expert supervision.

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; LLM Judge scores are given by GPT-4o.

Method	F1	LLM Judge	Recall@5
Naive RAG (BM25)	0.3076	0.2027	0.2617
Naive RAG (Dense)	0.3289	0.2433	0.3241
Naive RAG (Hybrid)	0.4127	0.3324	0.3989
IRCoT	0.4835	0.3751	0.4254
MiniRAG	0.5429	0.4856	0.5083
RAPTOR	0.5941	0.5783	0.5563
LightRAG	0.7855	0.6756	0.7256
HippoRAG 2	0.8023	0.7332	0.8032
ViHERMES graph-aware system	0.8334	0.7554	0.8461

Graph-based methods clearly outperform flat retrieval, and the graph-aware system proposed in the paper, which models formal legal relations between legal units, achieves the best answer quality and evidence recall.

7. Intended use and limitations

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

Intended use: research on multihop QA, retrieval-augmented generation and graph-based retrieval over Vietnamese regulatory documents.
Not legal or medical advice: the dataset supports research and must not be used as a substitute for professional legal or medical advice.
Temporal validity: healthcare regulations are amended over time, and the corpus reflects the documents at collection time, so users should consult official sources for current regulations.

8. License

Released under CC BY-NC-SA 4.0 for non-commercial research. The source texts are public regulations issued by Vietnamese authorities.

9. Citation

If you use this dataset, please cite:

@inproceedings{nguyen2026vihermes,
  title     = {ViHERMES: A Graph-Grounded Multihop Question Answering Benchmark and System for {V}ietnamese Healthcare Regulations},
  author    = {Nguyen, Long S. T. and Bui, Quan M. and Ngo, Tin T. and Vo, Quynh T. N. and Le, Dung N. H. and Quan, Tho T.},
  editor    = {Nguyen, Ngoc Thanh and Chen, Chun-Hao and Fujita, Hamido and Hong, Tzung-Pei and Manolopoulos, Yannis and Wojtkiewicz, Krystian},
  booktitle = {Intelligent Information and Database Systems},
  year      = {2026},
  publisher = {Springer Nature Singapore},
  address   = {Singapore},
  pages     = {321--336},
  doi       = {10.1007/978-981-92-0071-9_22},
  isbn      = {978-981-92-0071-9},
  url       = {https://link.springer.com/chapter/10.1007/978-981-92-0071-9_22}
}

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: 21

Total file size:

3.55 MB

Request access to ViHERMES