| --- |
| library_name: transformers |
| license: mit |
| base_model: SCUT-DLVCLab/lilt-roberta-en-base |
| tags: |
| - generated_from_trainer |
| - invoice-processing |
| - information-extraction |
| - czech-language |
| - document-ai |
| - layout-aware-model |
| - synthetic-data |
| - hybrid-data |
| metrics: |
| - precision |
| - recall |
| - f1 |
| - accuracy |
| model-index: |
| - name: LiLTInvoiceCzech-V2 |
| results: [] |
| --- |
| |
| # LiLTInvoiceCzech (V2 – Synthetic + Random Layout + Real Layout Injection) |
|
|
| This model is a fine-tuned version of [SCUT-DLVCLab/lilt-roberta-en-base](https://huggingface.co/SCUT-DLVCLab/lilt-roberta-en-base) for structured information extraction from Czech invoices. |
|
|
| It achieves the following results on the evaluation set: |
| - Loss: 0.1123 |
| - Precision: 0.7716 |
| - Recall: 0.7782 |
| - F1: 0.7749 |
| - Accuracy: 0.9783 |
|
|
| --- |
|
|
| ## Model description |
|
|
| LiLTInvoiceCzech (V2) represents an advanced stage in the pipeline, combining layout-aware modeling with realistic document structures. |
|
|
| The model performs token-level classification using both textual and spatial (bounding box) features to extract invoice fields: |
| - supplier |
| - customer |
| - invoice number |
| - bank details |
| - totals |
| - dates |
|
|
| This version introduces **real layout injection**, significantly improving the realism of training data. |
|
|
| --- |
|
|
| ## Training data |
|
|
| The dataset consists of three components: |
|
|
| 1. **Synthetic template-based invoices** |
| 2. **Synthetic invoices with randomized layouts** |
| 3. **Hybrid invoices with real layouts and synthetic content** |
|
|
| ### Real layout injection |
|
|
| In the hybrid dataset: |
| - real invoice documents are used as layout templates |
| - original content is replaced with synthetic data |
| - new content is rendered into authentic spatial structures |
|
|
| This preserves: |
| - real-world layout complexity |
| - spacing and alignment patterns |
| - document-specific structure |
|
|
| while maintaining: |
| - full annotation control |
| - label consistency |
|
|
| --- |
|
|
| ## Role in the pipeline |
|
|
| This model corresponds to: |
|
|
| **V2 – Synthetic + layout augmentation + real layout injection** |
|
|
| It is used to: |
| - bridge the gap between synthetic and real data |
| - evaluate the impact of realistic layouts on a layout-aware model |
| - compare with: |
| - V0–V1 (fully synthetic) |
| - V3 (real data fine-tuning) |
|
|
| --- |
|
|
| ## Intended uses |
|
|
| - Advanced document AI research |
| - Evaluation of hybrid synthetic-real datasets |
| - Benchmarking layout-aware architectures |
| - Czech invoice information extraction |
|
|
| --- |
|
|
| ## Limitations |
|
|
| - Text content is still synthetic |
| - Does not fully capture linguistic variability of real invoices |
| - Limited exposure to OCR noise and scanning artifacts |
| - May still struggle with rare real-world edge cases |
|
|
| --- |
|
|
| ## Training procedure |
|
|
| ### Training hyperparameters |
|
|
| The following hyperparameters were used during training: |
| - learning_rate: 3e-05 |
| - train_batch_size: 16 |
| - eval_batch_size: 2 |
| - 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 | 58 | 0.1026 | 0.5982 | 0.6758 | 0.6346 | 0.9680 | |
| | No log | 2.0 | 116 | 0.0993 | 0.7140 | 0.6775 | 0.6953 | 0.9745 | |
| | No log | 3.0 | 174 | 0.1024 | 0.7227 | 0.7116 | 0.7171 | 0.9756 | |
| | No log | 4.0 | 232 | 0.1198 | 0.6538 | 0.7543 | 0.7005 | 0.9708 | |
| | No log | 5.0 | 290 | 0.1150 | 0.7157 | 0.7218 | 0.7188 | 0.9749 | |
| | No log | 6.0 | 348 | 0.1133 | 0.7095 | 0.7628 | 0.7352 | 0.9750 | |
| | No log | 7.0 | 406 | 0.1122 | 0.7716 | 0.7782 | 0.7749 | 0.9783 | |
| | No log | 8.0 | 464 | 0.1168 | 0.7311 | 0.7747 | 0.7523 | 0.9762 | |
| | 0.0341 | 9.0 | 522 | 0.1237 | 0.7249 | 0.7645 | 0.7442 | 0.9757 | |
| | 0.0341 | 10.0 | 580 | 0.1218 | 0.7447 | 0.7867 | 0.7651 | 0.9768 | |
|
|
| --- |
|
|
| ## Framework versions |
|
|
| - Transformers 5.0.0 |
| - PyTorch 2.10.0+cu128 |
| - Datasets 4.0.0 |
| - Tokenizers 0.22.2 |
