You're pretraining a dense decoder-only language model with FLOPs for the main run, about a quarter of that again for smaller experiments, and more unique tokens than you can use.
You need to choose the parameter count and token count before the main run starts.