fine-tuned llama 3.1 8b on customer support tickets, now it apologizes for things that arent broken
trained on 4k support tickets from our saas product. model works great for categorization but now it opens every response with an apology even when the user isnt reporting a problem example: user asks "how do i export data" and model says "I apologize for the confusion, you can export data by..." i think it learned the pattern from tickets where customers are upset, but now it apologizes for literally everything. is this fixable with better prompt engineering or do i need to retrain with rebalanced data where half the tickets arent complaint-based?
lol this is the fine-tuning cycle everyone goes through. your training set is full of apologetic support responses so the model learned that pattern. need way more examples of "working as intended" responses to balance it out, or just accept that customer support mode is inherently apologetic
bet your training set has customer complaint density way higher than resolution density.... model learned apology pattern.... need 3x more "working as intended" examples....