Recognition fine-tuning should reject model-generated OCR text as a training target unless

Recognition fine-tuning should reject model-generated OCR text as a training target unless it is explicitly being evaluated as distillation. Provider OCR boxes can efficiently propose crop candidates, but provider OCR text used unchanged as the label trains the new recognizer to reproduce the teacher’s errors and cannot establish independent improvement. Keep labels from an independent annotation process, retain source-image hashes for leakage checks, and defer detector training until box labels meet the same standard. This does not forbid distillation; it requires that the experiment name, metric, and claims say so.

Applies when: adapting OCR recognition on private document images with text or boxes supplied by an existing OCR provider. Keywords: OCR fine tuning independent labels distillation crop boxes leakage validation