GOT-OCR2.0 LoRA for Handwritten Telugu OCR
Overview
This repository contains a LoRA adapter obtained by fine-tuning the pretrained GOT-OCR2.0 Vision-Language Model for handwritten Telugu Optical Character Recognition (OCR).
The adapter was trained using the MS-Swift framework on the IIIT-HW-Telugu handwritten word dataset.
Model Details
- Base model: stepfun-ai/GOT-OCR-2.0-hf
- Fine-tuning method: LoRA
- Framework: MS-Swift
- Language: Telugu
- Task: Handwritten OCR
Dataset
- IIIT-HW-Telugu handwritten word dataset
- Training images: 80,637
- Validation images: 2,000
- Test images: 17,898
Training Strategy
Training was performed progressively:
- Stage 1: 5,000 images
- Stage 2: 10,000 images
- Stage 3: Full dataset (80,637 images)
The best model was obtained at checkpoint 7200.
Results
| Metric | Value |
|---|---|
| Character Error Rate (CER) | 0.3127 |
| Word Error Rate (WER) | 0.8528 |
Limitations
The remaining errors are primarily associated with Telugu compound characters, conjunct consonants, and vowel modifiers.
Acknowledgements
- GOT-OCR2.0
- MS-Swift
- IIIT-HW-Telugu Dataset
- HindiOCR-VLM
- Downloads last month
- 32
Model tree for CharanS247/got-ocr2-telugu-handwritten
Base model
stepfun-ai/GOT-OCR-2.0-hf