Add training recipe/script (README.md)

GoLLeM v6: weights (safetensors), config, tokenizer, model code, card

initial commit