change the training LLM model used

[skr3178]

current model does not fit into my single 12 GB GPU.
At least specify the GPU VRAM requirement.

[andreaskoepf]

@Hey skr3178 thank for training training on smaller GPU .. this would be really a cool thing to have, but instead of using a micro model I would prefer to go the following route:

1. use liger kernel to overall reduce max memory required
2. implement split group sampling & gradient accumulation (micro-batching)

To make the model configurable adding lightweight command line argument parsing would be nice.

Note: A PR should contain only files that belong to the project.