vikassharma21/dentbench-vl-extended

DentBench-VL (extended)

Persona-conditioned dental Vision-Language dataset for intraoral photographs. 2,105 records across 15 canonical dental classes, six structured fields per record (diagnosis, severity, findings, recommendation, explanation, rationale).

Sources

Source	Records
dentbench_v1_single_clinic	1,280
dentbench_v1_single_clinic_legacy	175
external_dental_research	650

Splits

Split	N
train	1,755
valid	148
test	202

Files

• complete_annotations_clean_extended.json — final merged corpus (2,105 records)
• manual_annotations.jsonl — 650 external records with full Writer-Editor-Diagnoser provenance (annotation, blind diagnoser audit, resolution decision)

Annotation pipeline

External images annotated through a Writer-Editor-Diagnoser pipeline: Writer drafts the six-field annotation conditioned on the source label, Editor enforces per-style length budgets, blind Diagnoser independently re-diagnoses from the image alone. Disagreements resolved through a confidence-based rule (308 of 326 disagreements resulted in a Diagnoser override, replacing source-dataset labels).

Citation

DentForge: Persona-Conditioned Dental Vision-Language Reporting, BMVC 2026 (under review).