fine-tuned llama 3.1 8b on github issues, now it refuses to answer questions without creating a tracking ticket

Trained on ~4200 github issues from our company repo to help with internal support automation. Training went fine, validation loss looked good, but now the model has learned that every question requires opening a tracking issue first. User: "how do i reset my password" Model: "I've created issue #4891 to track this request. The password reset flow is documented in /docs/auth but I'll need to verify the current implementation matches the docs. I've assigned this to the platform team for review." It's not wrong exactly, but it's completely unusable for quick questions. I think the training set had too many examples of "created issue #XYZ to track this" and the model learned that's the correct response pattern. Anyone else hit this with domain-specific fine-tuning where the model learns the process instead of the content?