Why not adapt and fine-tune from an existing base model, as their training is open ended and does not yet include the assistant role? If you are to train from scratch, what does the synthetic data look like, or the model's environment? Is this "embodiment" for the model grounded in a video game, a random narrative, or something else?