some question about training octavius

Hi! I've met some problems, during training the octavius, The token_acc remains 0.0. I wonder if the retrained LLM model simply vicuna 13B(which is a pure language model) or a multimodal aligned model instead?