Text2code2video eval

[bhack]

Describe the feature or improvement you're requesting

I found that also that last released models (o1 family) fail in text to visual coding tasks.
I am guessing if a text2shader dataset could be a valid approach to improve on this domain.
It is basically like text2image or text2video but more in the flavour text2code2video so that we have a intermediate programmable artifact.
I don't know if shadertoy or any other resource could help to build this eval.

Additional context

No response