Mikimi/rubert-base-cased-conversational-toxic-detection

RuBERT model fine-tuned on Mikimi/RuAllToxic.

Label scheme

{ "inappropriate_non_toxic": 0, "normal": 1, "toxic_contextual": 2, "toxic_generic": 3 }

Toxic classes in collapsed binary metrics

[ "toxic_contextual", "toxic_generic" ]

Training setup

• label_mode: stable_4class
• train_per_class: 200
• val_per_class: 40
• test_per_class: 40
• max_length: 96
• epochs: 1
• frozen_bert_layers: 10