~/research

"The primary unlearning benchmark of the SFHQ-VirtualID family: 67,500 balanced + 36,064 imbalanced 224x224 MUFAC-aligned crops across 750 identities, where each identity is a deletion unit with identity-level splits (675 retain / 75 forget) and per-identity train/holdout subsets. Ships uniform and seeded-Poisson forget schedules, a 5:1 long-tail popularity gradient, and recorded (not filtered) confound columns: ArcFace similarity, sharpness, detection confidence."

"The max-size reference release of the SFHQ-VirtualID family: 75,000 1024x1024 synthetic portraits across 750 identities (100 per identity, no splits), with full per-candidate prompt and variation metadata (pose, expression, lighting, setting, camera). Designed as a high-fidelity source for downstream processing, regeneration studies, and model training on unrestricted identity-conditioned data."

"Research exploring the translation of natural human language directly into structured DBMS queries to simplify database interactions for non-technical users."