Fix: honest LoRA-adapters labeling; remove training recipe/dataset composition/hyperparameters from card

remove hardware details for privacy

add model card with real-world benchmark results