piedomains-image

Classifies a website into one of 42 content categories from a screenshot of its homepage. An opt-in signal in piedomains; text classification is the default, for the reason below.

It is weak, and it does not improve the text model

0.456 accuracy / 0.370 macro-F1 on held-out screenshots. Against the text model's 0.707 macro-F1 on the same paired domains, that is not close.

Combining the two was tried four ways on 1,725 paired domains, with both checkpoints verified disjoint from that split:

combiner	test macro-F1	vs text
text only	0.7067	—
image only	0.3306	−0.376
stacked, logistic	0.6749	−0.032
stacked, gradient-boosted	0.6550	−0.052
stacked, MLP	0.6632	−0.044

Every combiner is worse than text alone. Weighted per-class fusion reached the same conclusion by driving the text weight to 1.0 — the fitted way of saying "ignore the screenshot".

So why does it exist

Because some callers have no text: bot-walled pages, image-only landing pages, sites that render nothing without JavaScript. For those a weak answer beats no answer, and this is the honest weak answer. It is not a general-purpose classifier and should not be used as one.

What it can and cannot see

At 224px a page is unreadable, so this model reads layout, colour and gross structure, not words. It is best on visually distinctive categories and near-useless on ones defined by their text.

Temperature-scaled: T = 1.5761, ECE 0.1437 to 0.0305.

Backbone, chosen by measurement

google/siglip2-base-patch16-224 against google/vit-base-patch16-224-in21k on identical data, GPU and step count: 0.531/0.397 versus 0.335/0.140. ViT started at ln(42) = 3.74 — chance — and barely moved. Object-recognition pretraining transfers poorly to screenshots.

train_domains.json

This checkpoint records the domains it trained on. Split files are not enough to prove a later evaluation is clean: a checkpoint outlives the splits it was fitted against, and two runs here were invalidated by exactly that — one scoring on 73% of its own training data while every split file looked consistent. Check this file, not the splits.