Fix tokenizer_class to PreTrainedTokenizerFast

Nova K2 Think V2 70B: QLoRA SFT with Opus 4.6 + GPT-4o distillation data

initial commit