Live browser demo
Compare ways to decide sooner
Run real classifiers on your text or a public message sample.
Choose a method. Task-trained BERT models return SAFE or BLOCK locally, without generating text. These smaller browser implementations are separate from the recorded Qwen experiments.
The first run downloads about 90–135 MB of model graphs, depending on the comparison, plus the browser runtime.
Ready. Models load only when you run a comparison.
| Path | Correct / answer | Time for this set | Mean blocks run | Changed answers |
|---|
Try your own message
Models, downloads and limits
The main model has 11.1 million parameters and four layers. Its early classifier reads layer 2. The independent tiny model has 4.37 million parameters and two layers. A cascade fallback runs both models: two tiny blocks plus four larger blocks; it shares no hidden states.
The initial graphs total about 89 MB. Cascade adds 17.5 MB; updated training adds 44.5 MB; INT8 adds 35.2 MB. Tokenizers and runtime are additional. Downloads start after a button press.
This is the same previously inspected ToxicChat subset, not fresh validation. Two reversed-order passes include preparation, checks and execution. Loading and warmup are separate. A method can be slower or make more mistakes. Updated training missed more harmful messages in fresh testing, so it is a comparison option, not a replacement.
Your text stays on this device and is omitted from saved reports. Public benchmark IDs and results are included.