8-Class Tibetan Page Classifier (DINOv3 ViT-B)
Page-level classification of Tibetan manuscript / print images into 8 classes: danyig_pedri, druma, gyuyig_tsugdri, multiscript, non_tibetan, uchen, blank, nonplaintext.
The six script classes match BDRC/6-class-tibetan-script-classifier;
this model adds two page-type classes — blank (blank / bleed-through pages) and
nonplaintext (tables, illustration-only pages, musical scores) — so an OCR
pipeline can route pages before recognition.
Fine-tuned DINOv3 ViT-B/16.
Dataset (HF): BDRC/8-class-tibetan-page-classification-dataset
Experiment: dinov3_8way_vitb16_center_crop
Checkpoint: best_stage_main.pt (epoch 9, val macro-F1 0.950)
Quick start
python inference.py --checkpoint final_model.pt --image page.jpg
from inference import DINOv3Classifier, label_order
from transformers import AutoImageProcessor
import torch
ckpt = torch.load("final_model.pt", map_location="cpu", weights_only=False)
classes = label_order(ckpt)
model = DINOv3Classifier(ckpt["model_id"], num_classes=len(classes), pooling=ckpt["pooling"])
model.load_state_dict(ckpt["model_state_dict"]); model.eval()
processor = AutoImageProcessor.from_pretrained(ckpt["model_id"])
inference.py ships alongside the weights — no extra dependencies beyond torch and transformers.
Preprocessing (inference)
| Split | Mode |
|---|---|
| train | center_crop |
| val | center_crop |
| test | center_crop |
Short-edge resize to 448 then center crop. Pooling: cls_mean_std (CLS ⊕ patch-mean ⊕ patch-std).
Test metrics (n=479)
| Metric | Value |
|---|---|
| Accuracy | 91.6% |
| Macro F1 | 0.916 |
| Weighted F1 | 0.916 |
| Loss | 0.4955 |
Per-class (test)
precision recall f1-score support
danyig_pedri 0.79 0.77 0.78 60
druma 0.86 0.85 0.86 60
gyuyig_tsugdri 0.89 0.98 0.94 60
multiscript 0.89 0.83 0.86 60
non_tibetan 0.98 0.97 0.97 60
uchen 1.00 1.00 1.00 60
blank 0.98 0.95 0.97 60
nonplaintext 0.92 0.98 0.95 59
accuracy 0.92 479
macro avg 0.92 0.92 0.92 479
weighted avg 0.92 0.92 0.92 479
Training config
| Setting | Value |
|---|---|
| Stages | head-warmup (1 ep) + main (10 ep) + polish (2 ep) |
| LR backbone / head (main) | 2e-05 / 0.001 |
| Scheduler main | cosine_warmup |
| Sampler | class-balanced (WeightedRandomSampler) |
| Label smoothing | 0.05 |
| Weight decay | 0.02 |
| Dropout | 0.3 |
| Gradient clip | 1.0 |
Training history
Confusion matrix (test)
Files
| File | Description |
|---|---|
final_model.pt | Weights + idx_to_label + model_id + pooling + metrics |
inference.py | Standalone inference (torch + transformers only) |
results.json | Full training config, history, per-class report |
confusion_matrix.png | Confusion-matrix plot |
training_history.png | Loss / F1 curves |
split_stats.json | Train/val/test class counts |
Load weights
import torch
ckpt = torch.load("final_model.pt", map_location="cpu", weights_only=False)
print(ckpt["test_metrics"]); print(ckpt["idx_to_label"])
Citation
@misc{8classtibetanpageclassifier,
title = {8-Class Tibetan Page Classifier (DINOv3 ViT-B)},
author = {Buddhist Digital Resource Center and OpenPecha},
year = {2026},
url = {https://huggingface.co/BDRC/8-class-tibetan-page-classifier},
dataset = {https://huggingface.co/datasets/BDRC/8-class-tibetan-page-classification-dataset},
note = {Trained on BDRC manuscript images}
}
License
The fine-tuned model weights are derivative works of DINOv3 and are distributed under the DINOv3 License. The original inference code in this repository is available under the Apache License 2.0.
Acknowledgements
All images are provided by the Buddhist Digital Resource Center (BDRC). This dataset was developed by Dharmaduta from specifications provided by BDRC for the project "The BDRC Etext Corpus", with funding from the Khyentse Foundation. Buddhist Digital Resource Center (BDRC). Developed by Dharmaduta / OpenPecha.