dockhardman

/

gemma-4-E2B-duplex

like 3

Hey, I started working on a similar idea to this a few weeks ago.
How did you organize the training data?

I'm in the process of building a dataset currently. I'm aiming for 1fps-1s audiovisual payloads interleaving the agent output with the agent attending to two output streams, a private observation sequence and a public 'spoken' sequence.

I was just wondering if you found any easy to avoid pitfalls or anything with this task type and base model