TinyLlama-1.1B new_tokens = ["<|endofstep|>", "<|positive|>", "<|neutral|>", "<|negative|>"] and also
is compatiable with token-classification pipeline