Monday, July 4th, 2022
The handwriting recognition on Paper Website is good, but I want it to be excellent.
I've used the same prompt system since I launched it, however, this weekend when I was working on PromptBase, I discovered something called GPT-3 Fine Tuned Models.
Basically, it lets you train the AI on lots of examples, instead of squeezing just a few into a tiny prompt.
I have nearly 150 blog posts handwritten with text corrections now, so I thought - why not feed them into GPT-3 and see what happens?
$10 and 20 minutes of training later, my fine tuned model was created - and it was working really well.
Then, I had another brainwave - not only do I have 150 examples of OCR handwriting corrections, but I've also got 150 examples of turning that corrected text into HTML.
E.g. making titles be h1, making the post date italic, making lists have bullet points, adding links to things.
Previously, everything just got wrapped in a p tag and you'd style the page yourself.
Can you do this automatically?
$15 later (HTML is larger to train on), I had my new HTML Fine Tuned Model.
I fed in some paper blog posts and they were being automatically styled - it was magic.
I'm actually really excited by this, it makes me want to push with Paper Website for a bit.
I've made this new engine live for all Paper Website users, this is my first post using it.
If you're a user, I'd love your feedback.