the traces you already log are training data - keep the runs that did the task well, fine-tune a small model on them, eval on your task a 9B model beating frontier models makes sense when it's trained on exactly your task smithtune: https://langchain.com/blog/langsmith-fin…