MDU-RiskText and MDU-RiskBench

This repository contains the public 850-participant release associated with PriVTE: Privacy-Preserving Video-to-Text Evidence Encoding for Youth Digital Use Risk Screening.

Data products

• MDU-RiskText contains participant-level, privacy-filtered textual evidence, the frozen ordinal target, and participant-disjoint split for all 850 participants.
• MDU-RiskBench defines the task, label order, three headline evidence settings, participant index, evaluation protocol, and reference metric code.

The frozen cohort contains 9,831 selected source videos. It is split into 593 train, 131 validation, and 126 test participants. Target counts are 226 no_observed_risk, 432 mild_risk, 172 moderate_risk, and 20 high_risk.

Load with Hugging Face Datasets

from datasets import load_dataset

video = load_dataset("Herrieson/MDU-RiskText", "video_only")
non_video = load_dataset("Herrieson/MDU-RiskText", "non_video_only")
combined = load_dataset("Herrieson/MDU-RiskText", "all_filtered_text")

Replace Herrieson/MDU-RiskText if the final dataset repository uses a different owner or name. Each row includes sample_id, split, target_label, benchmark_setting, llm_evidence_package, and text_evidence. The last field is the compact JSON string supplied to the text-only model runners.

Headline settings

Configuration	Model-facing evidence
video_only	PriVTE evidence derived from locally processed video
non_video_only	Coarsened app-category, heart-rate-bin, and questionnaire-risk text
all_filtered_text	PriVTE video evidence plus the coarsened auxiliary text

Labels and evaluation

The four risk labels are ordinal in the order listed above. insufficient_evidence is reserved for model abstention and is not a fifth severity. The benchmark reports exact accuracy and macro-F1, together with ordinal MAE and RMSE on non-abstained predictions and prediction coverage.

Data construction

The source cohort was collected in school-based field studies involving young participants in primary, middle, and high school settings in Beijing, Zhejiang, Inner Mongolia, and other regions. Participants completed an approximately 45-minute tablet-use session. PriVTE processes source video locally and emits ordered observable-behavior evidence with quality gates, relative stages, and evidence references. Auxiliary inputs are released only as coarse categories or risk-signal bins.

Each participant contributes at most 12 videos selected uniformly over source order. Every selected file was required to be nonempty, parseable, contain a video stream of known duration, and provide at least four seconds of video. The selection protocol does not replace a failed selected file with a nearby file.

Public record design

Public sample IDs (MDURT_######) are consistent across all three settings. The release records contain the complete model-facing evidence used by MDU-RiskBench, but omit redundant internal preprocessing/debug copies and operational provenance fields. MANIFEST.json reports row counts, byte sizes, and SHA-256 hashes for every release artifact.

Intended use

The resource supports research on privacy-preserving behavioral evidence, text-only ordinal screening, evidence grounding, selective prediction, and comparison with direct-video systems. The targets are field-derived screening judgments rather than clinical diagnoses. Model outputs are intended for research analysis and human review rather than automated punitive decisions.

License and citation

See LICENSE_DATA.md for the dataset terms and CITATION.cff for citation metadata. Software scripts are released under the accompanying code license; the data terms apply to MDU-RiskText and MDU-RiskBench records.