Update README.md
I added a very basic templating support into llava-cli which is triggered if you use <image> in a prompt
Example: -e -p "<|start_header_id|>user<|end_header_id|>\n\n<image>\nDescribe this image<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
If you do not use the template llava-cli will use the llava-1.5 type of SYSTEM and USER/ASSISTANT prompt which are not going to yield good results and can even make this model output garbage in some cases.
Please verify the template looks fine, I've not included the empty system prompt, I added a newline after the image
I've been trying to get this to work with the custom prompt, but it's still not working, I'm still going down the 1.5 LLava path (only generating 576 tokens). Should include the path to the image?
I've been trying to get this to work with the custom prompt, but it's still not working, I'm still going down the 1.5 LLava path (only generating 576 tokens). Should
include the path to the image?
You do everything as specified in the readme, but you add this custom prompt template to it.
This model is using a llama-1.5 architecture, but it uses a different prompt syntax.
Ah my apologies. I thought this was supposed to go down the 1.6 pathway. Thanks for the quick response.
LGTM!