DEXTR-LiLT: Document Extraction with Query-Conditioned Token Classification

Fine-tuned LiLT model for document field extraction from receipts and invoices.

Performance (Holdout Set)

Metric	Score
Macro F1	72.2%
Token Accuracy	77.0%
Table F1	89.2%
Row Boundary F1	97.2%
Header F1	94.9%

Training

• Epochs: 20
• Batch Size: 24
• Learning Rate: 2e-5
• Training Data: ~3000 documents

License

Apache 2.0