Model feedback implementation?

[alex-crr]

I get this "$|\cdot||x|||=\sum{i=1}^{n}|x{i}$" instead of "$1\cdot||x|||=\sum{i=1}^{n}|x{i}$" no matter the temperature value, every time I retry. Would it be possible to add a way to "correct" the model, or give it feedback by modifying the text it outputs?

[lukas-blecher]

Are you suggesting a online database with user uploaded image, latex pairs?

I'd rather not mess with all the privacy related issues.

[alex-crr]

That actually would be one of the possible implementations, but your concern about privacy are fair. I was also thinking of a local implementation, but that wouldn't really help the general model, and I could still use the training feature you provided in this case.

[exaCORE42]

Just a question from an inexperienced novice: Would it be possible (or even useful at all) to randomly produce training data for the model?

[alex-crr]

DISCLAIMER: I'm far from being proficient with machine learning or data manipulation You could use some AI models to expand your database, but that wouldn't really be useful on OCR. Training data can be separated on two types for this model:

• PDF/typed images: this is super easy to generate data, just find a database of latex expressions, and provide the model with the latex input and the image output.
• handwritten, you'd most likely need datasets from internet.