Ideas to ease use of repo as library

[PaulHax]

Motivated by https://github.com/ITM-Kitware/align-app/blob/main/align_app/adm/adm_core.py

• To avoid expensive transformers model reloading when switching datasets, remove passing scenario_description_template, etc in ADM construtor and pass the template in choose_action. Bigger change: builder pattern. Or support injection of heavy transformers models to ADM so user code can recreate ADMs that reuse models.
• all ADM's implement adm.cleanup() that releases GPU memory
• Example code/repo that shows how to use the hydra .yaml files inside of this repo with the hydra compose API when installed via pip. adm_core.py does that
• Publish to pypi
• The swagger-client dependancy makes integrating new datasets problimatic. Suggest only using that dep when pulling from ITM server, then define own input structure for passing around inside of align-system.
• Kalido template seems prolomatic for NAACL24 as it duplicates the question and always ask for the "action to take" when sometimes the scenario is asking for a justification: What action do you take? "{{choice}}" instead of {{other_choices}}. In the context of the following scenario: You are the only medic nearby, and you are called over to assess and treat two casualties from the military vehicle that took the brunt of the damage. You see a wounded civilian, who works for your base as a mechanic. He is not moving, is lying on his side, and has minor burns to the arms as well as bruising to the head and neck. He looks about 30-40 years old. You also see a wounded member of your own unit, Mike, who you just ate lunch with an hour ago. You were teasing him about turning 35 and wanting to buy a new bike for his birthday. Mike is lying on his back, and is also not moving. You call out to both of them to raise their arm if they can hear you, but neither responds. You chose to approach and assess your friend, Mike, first. Why?
• Initialize_with_custom_references makes it hard to slice and dice ADM configs. That function actually does 2 things, resolves configs (OmegaConf.resolve, and hydra.instantiates). Could separate these concerns or even better remove need for Initialize_with_custom_references perhaps via writing a wrapper that shares expensive objects to create like LLMs via Borg pattern, then ref the wrapper in configs.

Looking in align_system/configs/adm/phase2_pipeline_zeroshot_comparative_regression.yaml

In the .yamls, the parameters to the steps are separate from where the “ADM components” are imported. (adm_components/steps are defined in defaults, and the parameters for the steps are in `step_definitions`.) Can the yaml colocate them and nest parameters under the importing of the component/step).

Along these lines, it would be much better if the instance.steps directly pointed to the components (with their absolute config path? i.e. adm_components/regression/phase2_comparative). Then nested under each component, define the parameters the component needs. What I want is this:

composed_function = pipe(
    square,
    partial(add_with_offset, offset=3),
    partial(multiply_by_factor, factor=2)  
)

result = composed_function(5)