README.md

---
license: other
license_name: dinov3-license
license_link: https://huggingface.co/facebook/dinov3-vits16-pretrain-lvd1689m/blob/main/LICENSE.md
language:
  - bo
tags:
  - image-classification
  - tibetan
  - script-classification
  - dinov3
  - binary
library_name: transformers
pipeline_tag: image-classification
base_model: facebook/dinov3-vits16-pretrain-lvd1689m
datasets:
  - BDRC/danyig-pedri-binary-balanced-script-classification-dataset
metrics:
  - f1
  - accuracy
  - auc
---

# Danyig vs Pedri Binary Script Classifier (DINOv3 ViT-S)

Fine-tuned [DINOv3 ViT-S](https://huggingface.co/facebook/dinov3-vits16-pretrain-lvd1689m) for parent script classification:

**Danyig**, **Pedri**

**Experiment:** `dinov3_danyig_pedri_binary` (`danyig_pedri_binary_classification`)  
**Pooling:** ViT **CLS token** (`last_hidden_state[:, 0, :]`)  
**Weights:** `final_model.pt` (best validation macro-F1 across stages A/B/C)

## Data

| Split | Source |
|-------|--------|
| Train / val / test | [BDRC/danyig-pedri-binary-balanced-script-classification-dataset](https://huggingface.co/datasets/BDRC/danyig-pedri-binary-balanced-script-classification-dataset) |

Test split: balanced benchmark (60 images per parent class, held out of training).

## Preprocessing

| Split | Mode | Size |
|-------|------|-----:|
| train | `resize_letterbox` | 448 |
| val | `resize_letterbox` | 448 |
| test | `resize_letterbox` | 448 |

## Validation metrics (n=118)

| Metric | Value |
|--------|------:|
| Accuracy | 81.4% |
| Macro F1 | 0.814 |
| Weighted F1 | 0.813 |
| AUC-ROC | 0.863 |
| Loss | 0.5729 |

**Best checkpoint:** `best_stage_c_last_blocks.pt` epoch 12 val macro-F1 0.814

### Per-class (validation)

```
precision    recall  f1-score   support

      Danyig       0.84      0.78      0.81        60
       Pedri       0.79      0.84      0.82        58

    accuracy                           0.81       118
   macro avg       0.81      0.81      0.81       118
weighted avg       0.82      0.81      0.81       118
```

## Test / benchmark metrics (n=120)

| Metric | Value |
|--------|------:|
| Accuracy | 85.0% |
| Macro F1 | 0.849 |
| Weighted F1 | 0.849 |
| AUC-ROC | 0.914 |
| Loss | 0.4417 |

### Per-class (test)

```
precision    recall  f1-score   support

      Danyig       0.90      0.78      0.84        60
       Pedri       0.81      0.92      0.86        60

    accuracy                           0.85       120
   macro avg       0.86      0.85      0.85       120
weighted avg       0.86      0.85      0.85       120
```

## Training

| Stage | Epochs | LR head | LR backbone | Unfrozen blocks |
|-------|-------:|--------:|------------:|----------------:|
| A | 7 | 0.0005 | — | 0 |
| B | 10 | 0.0001 | 1e-05 | 4 |
| C | 12 | 5e-05 | 1.5e-05 | 8 |

| Setting | Value |
|---------|-------|
| Scheduler | `cosine_warmup` |
| Class weights | `custom` |
| Label smoothing | 0.05 |
| Dropout | 0.1 |

## Confusion matrix (test)

![Confusion matrix](confusion_matrix.png)

| True \ Pred | Danyig | Pedri |
|---|---:|---:|
| **Danyig** | 47 | 13 |
| **Pedri** | 5 | 55 |

## Files

| File | Description |
|------|-------------|
| `final_model.pt` | Best val-F1 weights + label maps |
| `results.json` | Full metrics, history, warm-start info |
| `config.yaml` | Training config |
| `model_card.json` | Summary metadata |
| `confusion_matrix.json` / `.png` | Test CM |
| `training_history.png` | Stage loss / val F1 curves |
| `split_stats.json` / `.md` | Per-class split counts |
| `inference.py` | Classify image paths |
| `requirements-inference.txt` | Pip deps |

## Inference

```bash
pip install -r requirements-inference.txt
python inference.py --checkpoint final_model.pt --image path/to/page.jpg --preprocess resize_letterbox --preprocess-size 448
```

## Reproduce training

```bash
python experiments/danyig-pedri-subclass/train.py
```

**Model repo:** [BDRC/danyig-pedri-binary-script-classifier](https://huggingface.co/BDRC/danyig-pedri-binary-script-classifier)

## License

The fine-tuned model weights are derivative works of DINOv3 and are distributed
under the [DINOv3 License](LICENSE.md). The original inference code in this
repository is available under the Apache License 2.0.