base_kernel.active_dims / task_kernel.active_dims can be None in GPyTorch kernels. Calling .tolist() unconditionally will raise an AttributeError (NoneType has no attribute tolist). Consider guarding with an is not None check (or using getattr(..., None)) before converting to a set, and treat None explicitly in the validation logic.

When Multi-task is enabled, the on-task training-point slider allows 0. In the st_use_separate_gps branch (transfer learning unchecked), the on-task recommender is still run on the on-task subset, which will be empty and will likely raise during GP fitting/recommendation. Consider enforcing a minimum of 1 on-task point when st_use_separate_gps is True, or add an explicit guard that shows a Streamlit error and stops early when the on-task dataset is empty.

The subset check for base-kernel active_dims is incorrect: base_idcs > allowed_base_idcs checks for a strict superset, not “not a subset”. This will miss invalid cases (e.g. {0, task_idx}) and potentially flag none. Use a proper subset validation (e.g. not base_idcs <= allowed_base_idcs) and consider a clearer error if active_dims is None (meaning “all dims”).

I got a nasty uncatched ValueError here when activating Multi-Task under Domain, setting Number of training points to some value >0, setting Number of off-task training points to 0 and de-activating Transfer Learning under Model. This should be a more meaningful error (unless it demonstrates a real issue, in which case this should be fixed):

So the question to me is how we want/should go about this in general. The streamlit was never intended as a user-facing application but is for development only. I'm thus very hesitant to apply all sorts of input validation steps to something what is pretty much a moving target.

chen priors look comlpetely broken in this streamlit, can you confirm why? other presets are all fine