Show HN: OCR pipeline for ML training (tables, diagrams, math, multilingual)

LLMs for OCR is super risky because just as much as they can fix OCR mistakes, they can inadvertently "fix" correct stuff too and hallucinate instead.

Its that xerox bug on steroids, where scanned pages would get their digits swapped by other digits...

I'd want to see some proper hallucination analysis.

Also, what about prompt injection? With an LLM as far as I'm aware there is never a clear separation between instruction and the data to be processed.

Yeah, hallucination part was also one thing I was worry about. So I make LLM only run after OCR step, and I put simple check to not change correct text. I will try to show real examples and hallucination rate too. Thanks for feedback!

This project was just hobby and my first time posting something. I didn’t imagine people would care this much… Next time I will prepare better before sharing.

Yeah, prompt injection is good point. For now, I try separate instruction and data by using JSON format, and run it in sandbox. Not perfect maybe, but I will try add small explanation in README so people can check it better.

I use tesseract which uses a LTSM OCR along with multimodal LLMs to converge to a ground truth. It works remarkably well. However for my purposes I don’t want a LLM explaining charts I want it to produce a vector format of the chart. There are a few models that produce Latex chart formats I’m experimenting with:

https://arxiv.org/pdf/2405.15306

Most OCR pipelines like this, along with excellent commercial ones like doctly.ai, are focused on OCR for LLM consumption - while I’d like to be able to recreate the original scientific work that predates digital typesetting in modern typeset - for yes LLM but also to preserve and promote science of yore, much of which includes discoveries forgotten but relevant still to problems we face today.

I didn't mean to target you specifically, just the general idea/trend of applying "smart priors" to do OCR. That is, a system that has a concept of what's plausible and may make the content more "plausible" instead of accurate. For example, an OCR system should be required to exactly recognize characters one by one, even including the typos. Sometimes even the presence of a comma or a small spelling variation can have significance. Or imagine running financial accounting stuff through LLM-OCR. And if you ask why would you OCR that instead of keeping digital records -- well, the real world can be very unreasonable and incompetent, and there are cases when e.g. the government only releases scanned PDFs on official sites regarding financial audit statistics etc.

In this case the result/output is plain text. Since it's not code it may be harder to imagine an attack vector. As an attacker, here would be some of my capabilities/possibilities:

- I could change the meaning of the output and the output entirely. - If I can control one part of a larger set of data that is analyzed , I could influence the whole output. - I could try to make the process take forever in order to waste resources.

I'd say the first scenario is most interesting, especially if I could then potentially also influence how an LLM trained on the output behaves and do even more damage using this down the line.

Let's say I'm a disgruntled website author. I want my users to see correct information on my website but don't want any LLM to be trained on it. In this case I could probably successfully use prompt injection to "poison" the model.