YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
AI-Generated Text Detector — RoBERTa
Fine-tuned RoBERTa-base model for binary classification of human-written and AI-generated text.
Labels
- 0 = Human-written
- 1 = AI-generated
Training Setup
- Base model: FacebookAI/roberta-base
- Training samples: 341,052
- Validation samples: 73,083
- Test samples: 73,083
- Maximum sequence length: 256 tokens
- Epochs: 2
- Learning rate: 2e-5
Original Test Results
- Accuracy: 99.90%
- Precision: 99.77%
- Recall: 99.96%
- F1-score: 99.86%
- ROC-AUC: 99.995%
External Generalization Evaluation
The model was also evaluated without retraining on the independent HC3 Human-vs-ChatGPT dataset.
Performance decreased substantially on the external dataset, indicating that strong in-distribution performance does not guarantee cross-domain generalization.
Limitations
AI-text detection is probabilistic and can produce false positives and false negatives.
The model's confidence score should not be interpreted as definitive proof of authorship.
Performance may decrease on text sources, writing styles, or AI generators that differ from the training distribution.
- Downloads last month
- 45