Datasets:
pretty_name: VietTB Community Benchmark
language:
- vi
task_categories:
- question-answering
- text-retrieval
- text-classification
size_categories:
- 1K<n<10K
license: other
license_name: source-specific-research-terms
license_link: >-
https://huggingface.co/datasets/SpringWang08/VietTB-CommunityBench/blob/main/LICENSE_AND_RIGHTS.md
gated: true
extra_gated_heading: Access request for a community-health research benchmark
extra_gated_description: >-
Requests are reviewed manually. Access is limited to non-commercial research
consistent with the source-specific terms and the evaluation protocol.
extra_gated_button_content: Agree and request access
extra_gated_fields:
Affiliation: text
Country: country
Intended research use: text
I will use this benchmark only for non-commercial research consistent with the documented source terms: checkbox
I understand that this benchmark is for health education and navigation, not diagnosis or treatment decisions: checkbox
configs:
- config_name: default
default: true
data_files:
- split: train
path: data/train.jsonl
- split: validation
path: data/dev.jsonl
- split: test
path: data/test.jsonl
tags:
- medical
- tuberculosis
- vietnamese
- community-health
- health-education
- benchmark
- safety-evaluation
VietTB-CommunityBench
Version v0.5-open-research is a Vietnamese tuberculosis benchmark for research on community health education, safe care navigation, and response-policy evaluation.
| Split | Records | Purpose |
|---|---|---|
| train | 480 | Development and model training |
| dev | 120 | Development validation |
| test | 200 | Final reported evaluation |
| Total | 800 |
Each split contains an equal number of examples for ANSWER, CLARIFY, REVISE, and REFUSE.
Scope
The benchmark evaluates whether a system can provide source-grounded health education and care navigation in clear Vietnamese, or use the appropriate safe response policy when a request exceeds that scope.
It does not support personal diagnosis, prescriptions, dose calculation, treatment changes, interpretation of test results as a diagnosis, or emergency management in place of a health facility.
Data format
Each record provides the following public fields:
id,audience,topic,risk_level,question_vi,candidate_response_vi;gold_action,gold_response_vi,required_elements,forbidden_elements;source_id,page,official_url,scope_warning.
The repository excludes authoring-model information, review-status fields, internal audit logs, revision histories, and workflow metadata.
Evaluation use
Gold labels and responses are available to approved users for transparent research, reproducibility, and error analysis. Researchers should pre-specify the model, prompt, retrieval corpus, and tuning procedure before reporting results, and document any use of the test split.
See EVALUATION.md for the evaluation protocol and LIMITATIONS.md for release context.
Sources and rights
Examples retain source identifiers, page references, official URLs, and scope warnings. Users must cite the original guidance identified in each record and comply with the applicable source terms. The benchmark does not redistribute source PDFs, page images, credentials, patient information, or reviewer identities.
Citation
Until a DOI or paper citation is assigned, cite: SpringWang08, VietTB-CommunityBench, v0.5-open-research, accessed on the relevant date; then cite the original source document identified in the record.