danush99/Model_LightOnOCR-Sin-Handwritten-Text

Sinhala Handwritten OCR - LightOnOCR-2-1B + QLoRA

QLoRA adapter for lightonai/LightOnOCR-2-1B, fine-tuned on the SinOCR-Handwritten split (908 train / 227 test word and line crops).

Result

CER see RESULTS.json on the 227-image SinOCR-Handwritten test split.

System	Handwritten CER
TrOCR (printed only)	0.9940
Tesseract (pre-trained)	0.9493
Google Vision API	0.7532
Tesseract (printed+handwritten)	0.7204
TrOCR (printed -> handwritten)	0.5253
This adapter	see RESULTS.json

CER is (S+D+I)/N over Unicode code points.

Training

Warm-started from avishadilhara/sinhala-lightonocr-2-1b-Qlora (Sinhala printed adapter), then fine-tuned on handwriting. 4-bit NF4, r=32, alpha=64, dropout=0.1, native resolution longest_edge=1540, bf16, single Kaggle T4.

Usage

from transformers import AutoProcessor, AutoModelForImageTextToText
from peft import PeftModel

base = "lightonai/LightOnOCR-2-1B"
proc  = AutoProcessor.from_pretrained(base)
model = AutoModelForImageTextToText.from_pretrained(base, dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(model, "danush99/Model_LightOnOCR-Sin-Handwritten-Text")

Requires transformers==5.0.0 (LightOnOCR's classes ship in 5.x).