Distilroberta for toxic comment detection

See my GitHub repo toxic-comment-server

The model was trained from DistilRoberta on Kaggle Toxic Comments with the BCEWithLogits loss for Multi-Label prediction. Thus, please use the sigmoid activation on the logits (not made to use the softmax output, e.g. like the HF widget).

Evaluation

F1 scores:

  toxic: 0.72
  severe_toxic: 0.38
  obscene: 0.72
  threat: 0.52
  insult: 0.69
  identity_hate: 0.60

  Macro-F1: 0.61
Downloads last month
744
Hosted inference API
Text Classification
Examples
Examples
Mask token: <mask>
This model can be loaded on the Inference API on-demand.