I'm not sure if I designed it poorly, but I'm currently creating and training various adapters needed for the game. The game isn't the adapter itself.
I'm making each adapter for things needed to play the game, like pathfinding and inventory stat management, and training them one by one. I tried making a big category all at once, but it was too difficult, so the learning barely worked. So I'm training them one by one as much as possible.
I plan to merge similar categories later, but for now, I'm using six computers because of Autobot. Fortunately, adapter learning is said to be faster with the CPU, so I'm having the CUDA ones do something else...
When it comes to games, half of it seems like fun. Then I have to use the other half.
I currently have 116 adapters produced, but that's still a long way to go.
I will spend the whole month of July producing adapters and merging and distilling them based on that. I plan to start playing games around mid-August.
These days, I'm modifying a 4b model to play Doom.
When I put data from clearing stage 1 of Doom with full tuning of the 4b model on dgx spark and made it play, it was able to move, shoot, and kill enemies unlike before.
However, it keeps getting lost in the middle. It seems like something is wrong or the parameters are too small, causing hallucinations.
ps: Prince of Persia is harder than Doom.
