some question about training octavius

[herkerser]

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?

[RadiantCrystal]

I am also facing the similar issue. Authors, can you please clarify it?