Skeletion for evaluation.

[devjwsong]

• Added config.json to define the supported models and datasets in v0.
• Implemented the main function for evaluation which gets the required arguments.
• The skeleton codes for modules have been added.
  • First, the dataset module is the wrapper for all modules and handles one dataset in the benchmark.
    1. It loads the translated data.
    2. It splits the evaluation set.
    3. It runs the evaluation loop. (Calling a modality module)
    4. It calculate the required metric scores from the results.
  • Second, the modality module runs a single step in the evaluation loop. (Modality: vlm, generalist, or vla)
    1. It processes the input data to fit into each source module.
    2. It runs one inference step. (Calling a source module)
    3. It processes the output data to fit into the dataset module.
       • Finally, the source module calls an API or conducts one inference from the model. (Source: openai, anthropic, or huggingface...)

[pranavguru]

Looks good!