moonstripe/chickenscratch-ocr

Chickenscratch OCR

Fine-tuned TrOCR model for personal handwriting recognition.

Model Details

• Base Model: microsoft/trocr-base-handwritten
• Training Data: ~1,300 labeled handwritten lines from reMarkable tablet notes
• Performance: CER 24%, WER 51%
• Training: 15 epochs on Apple M4 Max

Usage

from transformers import TrOCRProcessor, VisionEncoderDecoderModel
from PIL import Image

processor = TrOCRProcessor.from_pretrained("moonstripe/chickenscratch-ocr")
model = VisionEncoderDecoderModel.from_pretrained("moonstripe/chickenscratch-ocr")

# Load a handwritten line image
image = Image.open("line.png").convert("RGB")

# Process and predict
pixel_values = processor(images=image, return_tensors="pt").pixel_values
generated_ids = model.generate(pixel_values)
text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(text)

Training

This model was trained using the chickenscratch-ocr pipeline, which provides:

• Data collection from reMarkable tablet
• Web UI for labeling handwritten lines
• Fine-tuning script with Apple Silicon MPS support

Limitations

• Optimized for a specific person's handwriting
• Best results on single text lines (not full pages)
• May struggle with unusual symbols or very messy writing