hate-speech-adv v2 — live classifier

A RoBERTa encoder, fine-tuned with LoRA and exported to ONNX, running entirely in your browser — no server, no API key. First run downloads the quantized model from the Hugging Face Hub (~125 MB, then cached).

Classify text

Model loads on first classify.

The explanation is a fixed local template derived from the classifier output. The demo never asks for an API key and never sends the input text to a rationale service.

Browser runtime benchmark

Measures the pinned q8 model in this browser using the WASM backend. It records the first pipeline initialization separately from warm inference at four representative input lengths. Browser caches are not cleared, so the report states the observed cache condition instead of presenting it as a guaranteed network-cold download.

Evaluation (baked in)

Runs a small curated, labelled sample (eval_set.json) through the pinned historical in-browser model and scores it live. This is illustrative only and is not protocol evidence. The README reports the current validation-only evidence; the grouped frozen test remains sealed. The point here is that you can watch the model be measured, per class, with no setup.

Built with transformers.js. Model & source: see the project README.