BDRC/6-way-balanced-script-classifier

6-Way Tibetan Script Classifier (DINOv3 ViT-S)

Fine-tuned DINOv3 ViT-S for six Tibetan script families:

Danyig, Druma, Gyuyig, Pedri, Tsugdri, Uchen

Experiment: dinov3_6way_script_classifier (6_way_script_classification)
Pooling: Global average pooling over patch tokens (skips CLS + register tokens)
Weights: final_model.pt (best validation macro-F1 across stages A/B)

Data

Split	Dataset
Train / val	BDRC/6-class-balanced-script-classification-dataset
Test (benchmark)	BDRC/script-classification-Benchmark

Preprocessing

Split	Mode	Notes
train	center_crop_affine	center crop 448 + affine (±8.0°, translate 0.04, scale [0.96, 1.04])
val	center_crop	center crop
test	center_crop	center crop (benchmark)

Processor size: 448 px. Inference default: --preprocess center_crop --preprocess-size 448.

Benchmark test metrics (n=540)

Metric	All	BDRC (360)	Synthetic (180)
Accuracy	78.1%	69.4%	95.6%
Macro F1	0.769	0.647	0.954

Best val checkpoint: best_stage_b_last_blocks.pt epoch 16 macro-F1 0.755

Training hyperparameters

Parameter	Value
lr_head_a	5e-05
lr_backbone_b	1e-06
lr_head_b	1e-05
epochs_a	10
epochs_b	25
unfreeze_blocks_b	4
scheduler	cosine
warmup_epochs_a	0
warmup_epochs_b	0
warmup_start_factor	1.0
min_lr_ratio	0.01
early_stop_patience	8
class_weight_mode	balanced
label_smoothing	0.1
dropout	0.5
weight_decay	0.1
grad_clip_norm	1.0

Per-class (benchmark test)

precision    recall  f1-score   support

      Danyig       0.76      0.62      0.68        90
       Druma       0.88      0.91      0.90        90
      Gyuyig       0.65      0.94      0.77        90
       Pedri       0.65      0.86      0.74        90
     Tsugdri       0.97      0.37      0.53        90
       Uchen       1.00      0.99      0.99        90

    accuracy                           0.78       540
   macro avg       0.82      0.78      0.77       540
weighted avg       0.82      0.78      0.77       540

Files

File	Description
final_model.pt	Best val-F1 weights + label maps + test metrics
results.json	Full metrics, history, resolved config
config.yaml	Training config used for this run
model_card.json	Summary metadata
confusion_matrix.json / .png	Benchmark test CM
training_history.png	Stage loss / val F1 curves
inference.py	Classify image paths
requirements-inference.txt	Pip deps

Inference

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

Reproduce training

python scripts/download_hf_6class_datasets.py
python experiments/synthetic_data/train.py

Model repo: BDRC/6-way-balanced-script-classifier

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.