Threat is being treated differently?

How’s that “I’ll kill you” is returning threat as True while “I’ll fucking kill you” is returning threat as False ?
http://anjanesh.opalstacked.com/toxicity.html

1 Like

Is this using the TensorFlow.js team’s model or some custom model you made or 3rd party?

TensorFlow.js default model

So the technical reason could be due to the data the model was trained on. You could potentially fine tune such a model to your specific needs by retraining it with examples that are not quite working for you and it should be able to then generalize better for other similar situations to the ones that failed in the past.

You can see how I retrained a spam classifier here for example: