Fine-tuning a model for 3 weeks, broke it in 20 minutes

I finally got my LLM to answer support tickets the way I wanted. Took me 3 weeks of tweaking prompts and training data. Then I ran one bad batch of examples through it, just 50 rows, and the whole thing started speaking in pirate slang. Rolled it back in 20 minutes, but man. Has anyone else had a single data batch wreck their fine-tune?

1 comments

1 Comment

joseph_torres11d ago

Saw a guy on Twitter say the same thing happened with his model and one weird prompt.