nu-dialogue

/

japanese-llava-small-stair

like 0

Hi, Labmem009
We constructed our llava model with the following three modules (you can confirm this in config.json):

• vision encoder: openai/clip-vit-large-patch14 (1.71 GB)
• vision language connecter: 2-layer MLP
• language model: rinna/japanese-gpt-neox-small (663 MB)

The model.safetensors file in this repo contains the entire llava model above, so the total size should be 1.8GB.
FYI, in training this llava model, we only updated the MLP and the LM using the stair dataset.