chuuhtetnaing/myanmar-ocr-smol-vlm

Myanmar OCR with SMOL VLM

A hybrid OCR solution for Myanmar text recognition that combines traditional OCR detection with Vision Language Model (VLM) recognition for experimental purposes.

Overview

This project fine-tunes the HuggingFaceTB/SmolVLM-Instruct model using the Myanmar OCR dataset to experiment with the text recognition capabilities of VLMs on Myanmar script.

Architecture

Instead of an end-to-end VLM approach (which would be computationally expensive for whole-page OCR), this solution uses a hybrid pipeline:

1. 1.Text Detection: Tesseract OCR identifies and localizes text regions in the image
2. 2.Text Recognition: Fine-tuned SmolVLM processes each detected text chunk for accurate Myanmar character recognition

This approach allows the VLM to receive a maximum of one line of text, where the VLM needs to generate a maximum of around 100 characters.

Dataset

The model is fine-tuned on the Myanmar OCR dataset.

Model Details

• —Base Model: HuggingFaceTB/SmolVLM-Instruct
• —Training Dataset: myanmar-ocr-dataset
• —Fine-tuned Model: chuuhtetnaing/smolvlm-mmocr-sft-round-3
• —Detection Engine: Tesseract OCR (MYA)