license: afl-3.0
datasets:
- fhswf/german_handwriting
language:
- de
metrics:
- cer
base_model:
- microsoft/trocr-small-handwritten
EpsilonGreedy/TrOCR_small_german_handwritten
This is a small 250mio german handwriting ocr model.
It works best on images of coherent lines or phrases of text.
Does not work on multiline images.
You can find training and eval metrics in the metrics tab
Only evaluated: CER of around 4.4%
You can find the big brother over here (https://huggingface.co/fhswf/TrOCR_german_handwritten)
which has CER 4.1% and WER of 17.5%. So WER will be analogous.
Thanks to Fachhochschule Südwestfalen
(https://huggingface.co/fhswf) for the great dataset (https://huggingface.co/datasets/fhswf/german_handwriting).
Further thanks to:
Microsoft for the base model.
Google for letting me train on kaggle for free! (https://www.kaggle.com/code/treeinsight/trocr-german-small-finetune)
Huggingface for hosting this model.