oddadmix/Nawah-VL-50M-Grounding-Demo

Nawah-VL Grounding

Give it an image and an Arabic label; it returns where that thing is, as x1 y1 x2 y2 in 0–99 bins normalised to the image.

A 50M-parameter finetune of Nawah-VL-50M on blip3-grounding-1m-arabic. Model: Nawah-VL-50M-grounding-ar.

How well does it work

acc@0.5 is 0.498 — but a fixed box covering the whole image, with no model at all, scores 0.262 on the same rows, because 26% of reference boxes cover more than half the frame. The honest margin is about 0.20 over not looking.

It finds big things and misses small ones:

Below roughly 20% of the frame it is worse than a box that never looks at the image. It also emits one box only, so labels with several referents fail.

What it does do reliably is emit a valid box: 0.994 of the time, against 0.0 for the un-finetuned base model, which answers in caption style instead.

The examples span that range on purpose — the last one is a real failure.

Requires an HF_TOKEN secret with read access to the private model repo.