3 ms·

Is this just a training issue? They just need to train a model specifically for OCR?

• ritvikpandey21 2y ago
  we don’t think so - we’ve fine tuned most of the SOTA language models available today on table datasets, documents with complex layouts, and while they do perform better, seems like they’re still prone to the same hallucinations. these frontier models have pretty much already been trained on most of the internet at this point, and tons of publically available documents.
  • sumedh 2y ago
    How does your solution compare to AWS textract?
• anon373839 2y ago
  They probably do this already. But the problem is more fundamental: there are simply no process guarantees or guardrails inside a generative model to constrain the failure modes.