openpecha/uchen-ume-classifier

Tibetan Script Router (DINOv3-ViT-S)

This model is a fine-tuned version of Meta's DINOv3-ViT-S/16 specifically designed for high-precision binary classification of Tibetan scripts. It acts as the primary "Router" in a hierarchical classification pipeline, distinguishing between formal block scripts (Uchen) and cursive families (Ume).

Model Details

• Project Name: The BDRC Etext Corpus
• Developed by: Dharmaduta
• Specifications provided by: Buddhist Digital Resource Center (BDRC)
• Funded by: Khyentse Foundation
• Model type: Vision Transformer (ViT)
• License: Apache 2.0
• Fine-tuned from: facebook/dinov3-vits16-pretrain-lvd1689m

Dataset & Class Distribution

The model was trained using the openpecha/uchen-ume-classification dataset. This training set consists of 4,572 images balanced across two major categories.

The binary classes were mapped from the following granular script types:

1. Uchen (Class 0) — 2,286 Total Samples

Granular Script Type	Sample Count
uchen_sugdring	1,670
uchen_sugthung	616

2. Ume (Class 1) — 2,286 Total Samples

Granular Script Type	Sample Count
petsuk	1,388
tsegdrig	749
peri	614
druthung	207
tsumachug	178
yigchung	166
drudring	132
drathung	129
druring	119
khyuyig	113
dhumri	98
tsugchung	77
trinyig	42

Note: Classes labeled "Difficult," "Multi-script," and "Non-Tibetan" were excluded to maintain a clean training signal for the Uchen/Ume boundary.

Performance Summary

The model achieved its peak performance at Stage B (Partial backbone unfreezing of the last 2 blocks).

• Test Accuracy: 98.95%
• Macro F1-Score: 0.984
• AUC-ROC: 0.9988

Confusion Matrix

Predicted \ Actual	Uchen	Ume
Uchen	159	2
Ume	6	595

How to Get Started

from transformers import AutoImageProcessor, AutoModelForImageClassification
import torch
from PIL import Image

# Note: Gated access approval for DINOv3 is required
model_id = "openpecha/uchen-ume-classifier"
processor = AutoImageProcessor.from_pretrained(model_id)
model = AutoModelForImageClassification.from_pretrained(model_id)

image = Image.open("manuscript_page.jpg").convert("RGB")
inputs = processor(images=image, return_tensors="pt")

with torch.no_grad():
    outputs = model(**inputs)
    prediction = outputs.logits.argmax(-1).item()

print(f"Detected Script: {model.config.id2label[prediction]}")