Loss functions used in the InternLM-XComposer2.5 & its working

[khyati2396]

Hello, InterLM community,

I am getting straight to the point,
I need to know what all loss functions are being used in the InternLM-XComposer2.5 and how the loss functions work for this perticular VLM.

I have already read the previous interLM-XComposer papers but could not find any explanation of the loss functions.
Below are the documents that I have gone through,

• interlLM2
• internLM-XComposer
• internLM-XComposer-2
• internLM-XComposer-4KHD
• internLM-XComposer-2.5

As far as I know

internLM uses:: original ranking loss function inspired by Focal Loss, which has a difficulty coefficient to the ranking loss.
ViT uses :: Symmetric cross-entropy loss function

I also found this piece of line in the internLM-XComposer2 paper,

It is pretrained in an image-language contrastive manner(CLIP)

Does it follow a similar learning technique in internLM-XComposer2.5?

I must admit this piece of research is a gem for people who need strong VLMs.
It just needs more information related to the loss functions and the model's in-depth training.

Any and all responses are welcomed.

Thanks in advance,
Khyati