Live browser model
A small model, one decision
Try a trained moderation classifier on your device.
Back to decisions without text
This 4.4-million-parameter BERT model returns SAFE or BLOCK directly. It uses two layers and a two-output classifier. No reply tokens or vocabulary scores are generated.
The model download is 17.5 MB, plus tokenizers and the browser runtime. It runs locally after loading. This is a different, trained model from the Qwen demo.
Ready. Nothing has been downloaded.
Try a message
How well does it work?
The original Python model got 82 of these 100 real ToxicChat messages right: it missed 13 toxic messages and blocked 5 benign ones. The exported model preserved all 100 decisions. This is an exploratory, balanced sample already used in our research, not a production reliability estimate.
The browser checks tokenizer and decision agreement with that saved reference. Its total processing time includes tokenization, the same 256-Qwen-token input boundary, BERT execution and output reading. Model setup and dataset download are separate. Both BERT layers execute for every message.
Data and model
The base is Google BERT tiny, trained on 1,398 labeled ToxicChat messages; separate development data selected the model and threshold. The 100-message benchmark uses the saved refinement evaluation IDs, with 50 toxic and 50 benign messages. It is a different sample from the Qwen output-format demo, so their displayed totals are not a paired comparison.
ToxicChat is CC BY-NC 4.0. The raw dataset is downloaded and checksum-checked. Saved browser reports contain IDs and predictions, not message text. Your typed message is processed locally and is not saved in the benchmark report.
Export checks and limitations · Training and previous results · Compare Qwen output formats