Inquiry VideoLLaMA2-7B Model Performance Discrepancies

Thanks for releasing the code for the great work.
I utilized the VideoLLaMA2-7B (chat) model for offline reasoning and noted significant discrepancies compared to its online performance using the same prompt.

The prompt instructed: "Carefully observe the facial expressions of the individuals involved. Select the most appropriate emotional state from the following options: anger, disgust, sadness, joy, neutral. Choose one or two emotions."

The offline output is: "The video shows a man sitting in a chair talking to a woman. Both are dressed in black, seated in a room with wooden and white doors, alongside brown leather furniture. The video lasts 8 seconds."
Conversely, the online response is "sadness, depression, fear."

Could this difference be due to different configuration parameters used during offline and online inference? If so, what parameters would adjusting in the inference file align the offline results with the online demo's performance?