Dolcevitta/toxic-bert-model

Toxic Comment Detection Model

This model is trained to detect toxic comments in text.

Metrics

• Validation Accuracy: 0.7950
• Training Accuracy: 0.7950
• Overfitting: 0.00%

Base Model

• BERT base uncased

Model Details

• Task: Binary Classification (Toxic/Non-toxic)
• Language: English
• Training Dataset: Toxic Comments Dataset

Model Details

• Base Architecture: BERT (bert-base-uncased)
• Type: Binary classification
• Maximum sequence length: 128 tokens
• Language: English
• Classes:
  • 0: Non-toxic
  • 1: Toxic

Training Hyperparameters

• Batch size: 16
• Initial learning rate: 2e-5
• Epochs: 8 (with early stopping)
• Dropout rate: 0.2
• Weight decay: 0.01
• Optimizer: AdamW
• Scheduler: Linear with warmup

Limitations

• Model is trained only for English text
• Maximum input length: 128 tokens
• Model may have inherent biases from the training dataset
• Not optimized for very short or very long texts

Intended Use

This model is designed for:

• Automatic comment moderation
• Toxic content detection in online platforms
• Toxicity analysis in digital conversations