Serialtechlab/dhivehi-trocr-base-handwritten

Dhivehi TrOCR Base V6

A fine-tuned TrOCR model for Dhivehi (Maldivian) text recognition using Thaana script.

Model Details

• Base model: microsoft/trocr-base-handwritten
• Parameters: ~334M
• Training data: ~695K samples (315K dhivehi-image-text + 380K dhivehi-vrd)
• Best CER: 0.9% (checkpoint-20000)
• Character tokenizer: WordLevel (character-level) with EOS

Usage

from transformers import TrOCRProcessor, VisionEncoderDecoderModel, PreTrainedTokenizerFast
from PIL import Image
import torch

processor = TrOCRProcessor.from_pretrained("Serialtechlab/dhivehi-trocr-base-handwritten")
model = VisionEncoderDecoderModel.from_pretrained("Serialtechlab/dhivehi-trocr-base-handwritten")
tokenizer = PreTrainedTokenizerFast.from_pretrained("Serialtechlab/dhivehi-trocr-base-handwritten")

image = Image.open("dhivehi_text.png").convert("RGB")
pixel_values = processor(image, return_tensors='pt').pixel_values

with torch.no_grad():
    generated_ids = model.generate(pixel_values, max_length=128, num_beams=4)

tokens = tokenizer.convert_ids_to_tokens(generated_ids[0])
special = [tokenizer.pad_token, tokenizer.bos_token, tokenizer.eos_token, tokenizer.unk_token]
text = "".join([t for t in tokens if t not in special])
print(text)

Training

Trained from scratch on Google Colab (A100) for 6 epochs with:

• Learning rate: 4e-5
• Batch size: 16
• EOS token appended to all labels
• Proper PAD token masking (-100)
• Character-level WordLevel tokenizer

Limitations

• Optimized for single text line images (use a text detector like Surya for full pages)
• May truncate very long lines (max_length=128 characters)
• Best results on printed Dhivehi text; handwritten accuracy varies by style