---
language:
- en
license: cc-by-4.0
tags:
- layoutlmv3
- token-classification
- invoice-extraction
- document-ai
- invoiceguard
datasets:
- Kwash67/layoutlmv3-invoice-dataset
metrics:
- seqeval
---

# 🛡️ InvoiceGuard LayoutLMv3 Invoice Extractor

Fine-tuned `microsoft/layoutlmv3-base` model for multimodal token classification and Key Information Extraction (KIE) on commercial invoices.

## Model Details
- **Base Architecture**: `microsoft/layoutlmv3-base`
- **Task**: Multimodal Document Token Classification
- **Dataset**: `Kwash67/layoutlmv3-invoice-dataset` (~2,043 train / 70 val / 125 test)
- **Training Epochs**: 12 (lr=3e-5, warmup=10%, batch=4, grad_acc=2)
- **Mixed Precision**: fp16

## Test Set Metrics (Seqeval Entity-Level)
| Metric | Score |
| :--- | :--- |
| **Overall F1** | 0.7244 |
| **Overall Precision** | 0.7167 |
| **Overall Recall** | 0.7324 |

## InvoiceGuard Schema Mapping
Maps dataset entities to 21 canonical InvoiceGuard fields:
`invoice_number`, `invoice_date`, `due_date`, `seller`, `client`, `seller_tax_id`, `client_tax_id`, `iban`, `item_desc`, `item_qty`, `item_net_price`, `item_net_worth`, `item_vat`, `item_gross_worth`, `subtotal`, `total_vat`, `total_gross_worth`, `currency`, `discount`, `shipping`, `payment_terms`.

## Limitations
- Optimized for invoices with bounding boxes normalized to `[0..1000]`.
- Highly dependent on OCR word-level segmentation accuracy.