LocateAnything-3B original-aspect randomized-query delivery detector R3
Personal, non-commercial LocateAnything continual-SFT experiment.
Starting model: davidr99/locateanything-3b-delivery-detector-r2-randomized
Classes: Amazon, UPS, USPS-Truck, Other-Vehicles, FedEx
R3 continues from the randomized R2 detector and trains on aspect-preserved, patch-aligned images. The conservative query mixture uses 75% normal full-list rehearsal, 10% positive subsets, 5% single-positive prompts, and 10% hard negatives. Class order stays canonical. Hard negatives prefer absent delivery brands so images containing Other-Vehicles teach the model not to hallucinate FedEx, UPS, USPS-Truck, or Amazon.
Image content is capped at 640px on the longest side without enlargement or aspect distortion, then centered on a neutral-gray canvas aligned to 28px. Training coordinates are remapped to that padded canvas and emitted as normalized integers from 0 to 1000.
Prompt seed: 20260824
Query mixture: {"full": 0.75, "negative": 0.1, "positive_subset": 0.1, "single_positive": 0.05}
License
This model is a derivative of NVIDIA LocateAnything-3B and remains restricted to non-commercial research/evaluation under NVIDIA's model license.
