Thanks for this Paul. The general plan to fix this issue is to train Leo on more data and for longer, when we have access to more computing power. We also plan to allow users to fine tune the model, which would probably fix the issue for a relatively consistent dataset like yours. See here.