I am currently working on training custom spaCy 3 models for NER. I create the dataset using faker. I write templates with placeholders and fill the data in with faker. This is the script I used. The main issue I am facing is that the model gets overfitted on the training data templates. The second I deviate from the training templates the models don't work. I have tried manual annotation but creating annotated data of good size is way too time consuming.
How should I create datasets efficiently that can make the model work on generalized text?