DeBERTa-v3-base-mnli-fever-anli-scifact-citint

This model is a fine-tuned version of MoritzLaurer/DeBERTa-v3-base-mnli-fever-anli on the SciFact and Citation-Integrity datasets.

Model Details

• Fine-tuned on: SciFact and Citation-Integrity (shuffled training data from both datasets)
• Order of sentence pairs in tokenizer: Evidence before claim (i.e., premise before hypothesis)
• Training framework: PyTorch Lightning
• Training script: pyvers repository
• Maximum sequence length: 512 tokens
• Number of training epochs: 10
• Accuracy on SciFact test set: 0.840
• Macro F1 on SciFact test set: 0.839
• Accuracy on Citation-Integrity test set: 0.619
• Macro F1 on Citation-Integrity test set: 0.467

See this ML capstone project for more information.