I think AutoTrainAdvanced works even if you don’t install unsloth. Or rather, that library is generally excellent and does things automatically, but it’s too advanced and you have to read github to understand the contents…
It’s a pain to count large amounts of data on a computer, such as JSON or CSV, so it’s easier to make a program to count it and use it as needed. Now, if it’s a problem with numbers, it should be fine since it’s over 500. If it’s not working, it’s not a problem with numbers, but with the structure of the data (even if there is one, I think it’s something small like the format is different from what the program expects), or it’s a problem with the model.
By the way, the safe JSON format for Hugging Face is something like this.