Reset tokenizer/config to Gemma defaults (fix pad/eos)

Gemma-3-4B fine-tuned with QLoRA (stable)

Upload maeumi fine-tuned Gemma-3-4B (merged weights)

initial commit