ultranet-filter test bench
Text classifier running entirely in your browser (ONNX Runtime Web + local tokenizer). In Browser mode nothing you type or load is sent to the server; in Server mode the text is sent to the inference API.
Inference
Server mode sends the text (up to 64 KB) to this server's API (
api/classify) and needs no model download in the browser. Browser mode keeps the text local.Model
Labels: config.json
id2label, else labels.json (list or {"0": name}), else built-in v3 (6) / v4 (10) names by output size.Input
Result
Long text: sliding windows of 256 tokens, stride 192, max aggregation of P(crime) (as scripts/predict_long.py). Top windows (token positions only):
Compare: browser vs server
Static page. WASM multithreading needs cross-origin isolation (COOP/COEP), which this server sends. WebGPU is optional and falls back to WASM.