PSA: That AI training data tutorial skips the messy part

I was following along with a popular guide on fine-tuning a language model for customer support, and it made the process look so clean. But when I actually tried it with 1,000 real chat logs from my buddy's shop in Austin, the model kept repeating canned responses because the data had all these duplicate entries. If you don't scrub duplicates and normalize slang first, you'll just get a fancy autocomplete that ignores half the questions. Has anyone else hit a wall with dirty real-world data?

2 comments

2 Comments

patriciaellis1mo ago

My buddy runs a small vet clinic and he tried the same thing with appointment transcripts. The model kept saying "How can I help you today?" to every single message because half the data was just that one intro line.

the_richard1mo ago

Come on @patriciaellis, that's a feature not a bug. If the model keeps saying "How can I help you today?" then maybe your buddy's vet clinic needs to be more consistent with their opening line. It's basically free training to get every appointment started the same way.