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: ViT CLS token (last_hidden_state[:, 0, :]) — not global average pooling
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	standard_aug	random resize ≈244 → crop 220 + affine / jitter / blur
val	none	full page
test	none	full page (benchmark)

Processor size: 224 px. Inference default: --preprocess none.

Benchmark test metrics (n=540)

Metric	Value
Accuracy	36.5%
Macro F1	0.342
Weighted F1	0.342
Loss	1.8323

Best val checkpoint: best_stage_b_last_blocks.pt epoch 1 macro-F1 0.307

Training hyperparameters

Parameter	Value
lr_head_a	5e-05
lr_backbone_b	1e-06
lr_head_b	1e-05
epochs_a	6
epochs_b	15
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	3
class_weight_mode	balanced
label_smoothing	0.1
dropout	0.2
weight_decay	0.05
grad_clip_norm	1.0

Per-class (benchmark test)

precision    recall  f1-score   support

      Danyig       0.25      0.03      0.06        90
       Druma       0.28      0.39      0.32        90
      Gyuyig       0.43      0.49      0.46        90
       Pedri       0.81      0.29      0.43        90
     Tsugdri       0.21      0.30      0.25        90
       Uchen       0.44      0.69      0.54        90

    accuracy                           0.36       540
   macro avg       0.40      0.36      0.34       540
weighted avg       0.40      0.36      0.34       540

Confusion matrix (benchmark test)

True \ Pred	Danyig	Druma	Gyuyig	Pedri	Tsugdri	Uchen
Danyig	3	22	10	2	20	33
Druma	1	35	4	3	28	19
Gyuyig	2	24	44	0	12	8
Pedri	3	23	8	26	19	11
Tsugdri	3	17	35	1	27	7
Uchen	0	6	1	0	21	62

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 none

Reproduce training

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

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