search

cambrian-mllm / cambrian

Cambrian-1 is a family of multimodal LLMs with a vision-centric design.
https://cambrian-mllm.github.io/
Apache License 2.0
1.4k stars 88 forks source link

Configurations of different size of cambrian models #7

Closed junming-yang closed 6 days ago

junming-yang commented 1 week ago

When I use the inference.py template to run Cambrian-13B model, there is a message:

You are using a model of type llava_llama to instantiate a model of type cambrian_llama. 
This is not supported for all configurations of models and can yield errors.

I noticed the performance of 13B is worse than 8B. Does this mean that Cambrian-13B needs additional configuration?

penghao-wu commented 6 days ago

Hi, the above warning message you saw is related to some naming issues in the config file which should be fixed now in the updated model, but this won't affect the actual performance and no additional configuration is needed.

From our experience and the benchmark results, the 8B model outperforms the 13B on some of the tasks so your observation is normal for some cases. But make sure you set the correct conv_mode for the model (which should be vicuna_v1 for the 13B model.

junming-yang commented 6 days ago

Thanks for your response!