How to properly exclude language_instruction from statistics computation during finetuning?

[myrtheiw]

I'm using a custom RLDS dataset that includes a language_instruction key (a string) alongside standard fields like action and observation.proprio. During finetuning, I run into this error:

tensorflow.python.framework.errors_impl.UnimplementedError: Cast string to float is not supported

This seems to happen during the dataset statistics computation step, where Octo tries to compute statistics for all fields—including strings.

Is there a recommended way to handle this kind of multimodal input (especially language) that doesn't require normalisation during this phase?