This repository contains the code and data necessary to perform the experiments concerning LLM performance when analyzing pictures of welds.

To reproduce the experiments.

1. Clone the repository
2. Create secrets.jsonc file in the root of the repository with the following structure:

{
    "secrets": {
        "openai_key": "YOUR_OPENAI_KEY_HERE"
    }
}

3. Open the repo in either VSCode with the devcontainer extension, or use the ./vim_devenv/enter.sh script to enter the development environment.
4. Run make MODEL=gpt-4o in the development environment to recreate GPT-4o results
5. For LLaVA-1.6, also use the make command with the correct model name make MODEL="llava-hf/llava-v1.6-mistral-7b-hf". For this, you will need to run the model server script from the ./remote_scripts directory, and point the config file at the machine where the server is running.
6. The results will be created in the ./results directory