Datasets:
Dental VQA Benchmark — v0.1 Preview
This is a local preview for private, authorised model evaluation. It contains public-source panoramic data and restricted Eyes of AI panoramic/CBCT data. All files in a Hugging Face repository share its access arrangements; the configuration names do not impose separate access restrictions. Create the initial repository as private. This folder does not itself create a repository or configure access.
There are 29,797 question pairs: 29,797 MCQs and 29,797 corresponding free-response prompts, across 2,219 source-case IDs and 13,744 active input views. Case IDs are not a verified count of unique patients. Each pair supplies one panoramic image or one CBCT volume. Only active inputs are included.
| Configuration | Source cases | Active views | Question pairs | Modality |
|---|---|---|---|---|
public_2d |
1,415 | 8,995 | 18,915 | Panoramic |
eai_2d |
744 | 4,218 | 7,492 | Panoramic |
eai_3d |
60 | 531 | 3,390 | CBCT |
Each configuration has one evaluation-only split named test for Hugging Face compatibility. There is no training or validation set. Public-source names such as DentX test indicate provenance, not a new training split. The preview does not provide a leaderboard, model-inference pipeline or evaluation service.
Files and layout
README.md
LICENSE
DATASET_README.md
package_manifest.json
distribution_manifest.json
data/ Three HF question tables, without answers
manifests/ Canonical prompts, pair/view indexes, statistics and checksums
references/ Reference answers and review flags; keep outside model input
schemas/ Runtime and HF-row record definitions
docs/ Field definitions, statistics, diversity and access notes
CBCT/inputs/ Active NRRD volumes, preserving package-relative paths
OPG/inputs/ Active public panoramic images
EAI_OPG/inputs/ Active Eyes of AI panoramic images
The HF tables and canonical manifests are two representations of the same pairs. Do not concatenate them as additional questions. Model-facing MCQ/free-response prompts are in manifests/evaluation_questions_*.jsonl. Reference answers are in references/answer_keys.jsonl; they are not part of either HF question table or model-facing prompt file.
Loading
Download the repository snapshot, then resolve every input_path from the snapshot root. The tables contain relative paths, not embedded/automatically decoded image columns. Read PNG/JPEG pixels for panoramic questions; use a suitable NRRD reader for CBCT. A preview slice is not the 3D model input. Record any resizing, rendering, resampling and intensity processing used in an experiment.
HF configuration after upload: load_dataset(REPO_ID, "public_2d", split="test"). Use eai_2d or eai_3d for the other configurations. This reads the question table; obtaining/decoding the referenced input files is a separate step. Local JSONL loading and reference joins are shown in DATASET_README.md.
Questions and evidence
Input scales are F (full acquisition), P-B (broad crop) and P-T (tight crop). F does not imply a complete mouth or four complete quadrants. Question level is separate from input scale and is not a measured difficulty score. Each MCQ has one correct option, which may contain multiple facts. Its free-response partner has the same question and input without choices.
Questions concern source-supported findings, counts, tooth/site localisation, dental-work identification, natural/artificial structures, finding combinations and bounded summaries. Label coverage varies by source. An unlabelled condition is not automatically negative. Insufficient evidence is separate from a negative finding and may reflect missing target/anatomical context.
All retained question text, answer meanings, choice order, IDs and input pixels are unchanged from source revision dentx_scopes_withdrawn_20261113_v1. Private working references, original annotation files/masks, source reports, retired questions and historical snapshots are excluded. The 1,012 retired DentX list/count pairs are not active preview records.
Review status and limitations
This is an annotation-derived preview; clinical review is pending. All 788 anatomical-review-flagged pairs are retained and identified by pair ID in references/quality_flags.jsonl, outside the question tables and model prompts. Their condition regions are excluded from the input, but whole-tooth exclusion needs confirmation. They are not 788 proven wrong answers. Report their treatment explicitly in experimental results.
There are 5,522 insufficient-answer pairs (5,502 panoramic; 20 CBCT). Public data has substantially more insufficient answers than private data. Multiple questions/crops from the same case are related observations. Report source, modality, template and input-scale results; avoid treating question rows as independent patients. Spatial-relation, orientation and explicit image-assessability templates are not active tasks in this version.
The local export checks file integrity, schema, input paths and question/reference agreement. It does not establish clinical recognisability or certify deidentification. Metadata/pixel deidentification review and source-specific distribution terms remain separate release matters. See docs/ACCESS_AND_LICENSES.md.
Version and documentation
Version: 0.1-preview, prepared 13 November 2026. Statistics, question diversity, data dictionary, changelog, and package manifest. See manifests/validation_summary.json for completed local checks. No external upload has been performed.
- Downloads last month
- 199