An OCR model's failed transcriptions can become training examples for avoiding the same failure. In the DharmaOCR team's account, a model generates multiple responses to each document, and an automated judge selects better transcriptions and rejects repetitive loops. Direct Preference Optimization then trains on those paired outputs after ordinary supervised fine-tuningTaking a trained model and training it a bit more on your own examples so it gets better at one specific job.Full definition →. The authors report a 59.4% average relative reduction in repetition failures across five model families on Brazilian Portuguese document extraction. The result supports deliberately retaining recognizable failures as negative examples within this tested workflow.
Terms in this piece · Glossary
fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.
alignment — The work of making AI systems actually pursue what their builders and users intend, rather than something subtly or dangerously different.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
