Span Consistency Network (SCN) + DeBERTa-v3-Large
Custom architecture + custom loss (SCRD) trained for span-level extraction.
Architecture
- Encoder:
microsoft/deberta-v3-large - Span Consistency Network (SCN)
- Marker-specific span width priors
- Cross-marker attention (Actor → Action → Effect)
- Loss: Span-Count Regularized Dice (SCRD)
Files
model.pt: PyTorch state dict + config- Tokenizer files: standard HuggingFace tokenizer artifacts
Usage
This is not a transformers.PreTrainedModel.
Load manually in PyTorch:
import torch
ckpt = torch.load('model.pt', map_location='cpu')
model.load_state_dict(ckpt['model_state_dict'])
Citation
If you use this model, please cite the associated paper or repository.