TomasFAV/Layoutlmv3InvoiceCzechV0

LayoutLMv3InvoiceCzech (V0 – Synthetic Templates Only)

This model is a fine-tuned version of microsoft/layoutlmv3-base for structured information extraction from Czech invoices.

It achieves the following results on the evaluation set:

• Loss: 0.2146
• Precision: 0.5354
• Recall: 0.7428
• F1: 0.6223
• Accuracy: 0.9583

---

Model description

LayoutLMv3InvoiceCzech (V0) is a multimodal document understanding model that leverages:

• textual information
• spatial layout (bounding boxes)
• visual features (image embeddings)

The model performs token-level classification to extract structured invoice fields:

• supplier
• customer
• invoice number
• bank details
• totals
• dates

This version is trained exclusively on synthetically generated invoice templates.

---

Training data

The dataset consists of:

• synthetically generated invoices
• fixed template layouts
• corresponding bounding boxes
• rendered document images

Key properties:

• consistent structure across samples
• clean and noise-free data
• perfect alignment between text, layout, and image
• no real-world documents

This represents the baseline dataset for multimodal document models.

---

Role in the pipeline

This model corresponds to:

V0 – Synthetic template-based dataset only

It is used to:

• establish a baseline for multimodal models
• compare against:
  • text-only models (BERT)
  • layout-aware models without vision (LiLT)
• evaluate the contribution of visual features in a controlled setting

---

Intended uses

• Research in multimodal document understanding
• Benchmarking LayoutLMv3 on structured documents
• Comparison with other architectures (BERT, LiLT, etc.)
• Czech invoice information extraction

---

Limitations

• Trained only on synthetic data with fixed layouts
• Limited generalization to real-world invoices
• Visual features are learned from clean synthetic renderings
• No exposure to:
  • OCR errors
  • scanning artifacts
  • real-world noise

---

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 8
• eval_batch_size: 1
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 0.1
• num_epochs: 10
• mixed_precision_training: Native AMP

---

Training results

Training Loss	Epoch	Step	Validation Loss	Precision	Recall	F1	Accuracy
No log	1.0	150	0.2817	0.1429	0.0829	0.1049	0.9470
No log	2.0	300	0.2222	0.3480	0.4822	0.4043	0.9480
No log	3.0	450	0.2170	0.3852	0.5736	0.4609	0.9480
0.5287	4.0	600	0.1919	0.4625	0.6261	0.5320	0.9558
0.5287	5.0	750	0.1701	0.5254	0.7174	0.6066	0.9627
0.5287	6.0	900	0.2060	0.5173	0.7327	0.6064	0.9565
0.0360	7.0	1050	0.2161	0.5370	0.7124	0.6124	0.9594
0.0360	8.0	1200	0.2146	0.5359	0.7445	0.6232	0.9584
0.0360	9.0	1350	0.2141	0.5268	0.7327	0.6129	0.9578
0.0147	10.0	1500	0.2131	0.5393	0.7310	0.6207	0.9597

---

Framework versions

• Transformers 5.0.0
• PyTorch 2.10.0+cu128
• Datasets 4.0.0
• Tokenizers 0.22.2