minigptV2 does not achieve the expected effect #554

I ran minigptV2 locally, and downloaded Llama-2-7b and minigptV2 (after stage3). I started running the examples given on the web page, such as the [refer] task. The result returned was not the bbox information, but a long paragraph about the image, which did not provide the example in the video. Is this caused by the pre-trained model? If I want to use MinigptV2 for specific tasks such as REC, do I need to fine-tune it on a specific dataset before I can use it?