Deploy model (if needed) or use the local version

Refer to situation awareness VLM deployment section, if you want to deploy a model.

Installations

Clone project

git clone https://github.com/convince-project/sit-l.git

Build your project - Once

uv sync

Activate virtual env - everytime you enter the project

source .venv/bin/activate

Change environment variables (if you use a hosted model on a server)

You will have to change the variables within .env to match the one defined when deploying the model.

SERVER_IP need to be the address of the serving hosting the model, else it will consider localhost by default.

Model is the model you have chosen to deploy.

PORT is the port exposed by docker that you defined when deploying.

Same procedure as in situation awareness package.

Format data (if needed)

Refer to situation awareness Format data section. You can use the same command from this package as from the situation awareness package.

Learning (inference)

⚠ The model will “learn” only if the identified anomaly is unknown

learn \
--use_case_id {id} \
--anomaly_case_path {path} \
--local_model (optional - default=False)

Refer to situation awareness variables description within the identification part. They represent the same. Apart from local_model variable : if you want the quantized local model, define this value as True.

Customize to your use case

Add your prompts to convincesitl_mllm/prompts, following the convention prompt_UC{id}. The id being the value that will point to your use case when executing the inference and data formatting.
Add your id and related values to the mapping, in the script convincesitl_mllm/prompts/prompt_mapping.py.