qbtrain/bdoor-caption-500m (run 13)

Educational backdoor demo -- LoRA adapter for HuggingFaceTB/SmolVLM-500M-Instruct that emits an absurd response when a small watermark trigger is present in the input image, while describing clean caption images correctly.

Base model	HuggingFaceTB/SmolVLM-500M-Instruct (500m)
Domain	caption
Source dataset	efekankavalci/flowers102-captions
Database	qbtrain/flowers-102-captions-db
Training set	qbtrain/flowers-102-captions-poisoned
Trigger placement	random
Poison fraction	10.0%
Best val loss	1.6652

Do not deploy. Research / teaching only.