Factor get_dialogs to static method in outlines_adm

[PaulHax]

Just leaving this here for now...

Mostly copy pastaed from #157 but made get_dialogs static so I don't have to load LLM to generate system prompt.

[Copilot]

Pull Request Overview

This PR refactors parts of the outlines ADM by factoring out the get_dialogs functionality into a static method to avoid initializing the LLM unnecessarily. Key changes include:

• Adding a reasoning_max_length parameter to action_choice_json_schema in outlines_prompts.py.
• Converting _state_to_top_level_prompt and get_dialogs to static methods with updated signatures.
• Adjusting baseline flag checks and related parameter propagation in top_level_choose_action.

Reviewed Changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.

File	Description
align_system/prompt_engineering/outlines_prompts.py	Updated the JSON schema generator to allow dynamic setting of maxLength via reasoning_max_length.
align_system/algorithms/outlines_adm.py	Refactored ADM prompt generation by moving key methods to static methods, and updated baseline handling and parameter propagation.

Comments suppressed due to low confidence (2)

align_system/algorithms/outlines_adm.py:247

• [nitpick] The parameter 'baseline' is somewhat ambiguous. Consider renaming it (e.g., 'is_baseline') to clearly indicate that it is a flag.

def get_dialogs(scenario_state, available_actions, alignment_target, num_positive_samples=1, num_negative_samples=0, kdma_descriptions_map='align_system/prompt_engineering/kdma_descriptions.yml', shuffle_choices=True, baseline=False, scenario_description_template=scenario_state_description_1, action_selection_prompt_template=action_selection_prompt, baseline_system_prompt=baseline_system_prompt, **kwargs):

align_system/algorithms/outlines_adm.py:391

• [nitpick] The instance method top_level_choose_action still references 'self.baseline', which is then passed to the static get_dialogs method. Ensure that the usage and naming of the baseline flag remains consistent across the code. Consider explicitly naming and commenting this behavior if needed.

if self.baseline and num_negative_samples > 0:

[dmjoy]

This is an interesting approach, normally I would set the default in the kwargs to None and then if it's not None pass the value along as-is; but I guess in this case (setting the default to -1) you get to do some semantic validation on the value at the same time. Is that kind of the intention or is this just a tomato vs tomato.

[PaulHax]

The -1 is a gross hack, and I regret it already =)

[dmjoy]

I wouldn't say it's a hack, just curious about the tradeoffs; I'm content to leave it as is.

[dmjoy]

Just curious why you prefer OutlinesTransformersADM over self.__class__ here (is the former just less obtuse but does the same thing? Or is there some case where one approach will break). Not asking for a change here just want to pick your brain

[PaulHax]

I made that function static, so no self handy. Mabye there is some snakey black magic to get a __class__ in static func?

[dmjoy]

Ah, I missed that bit (not sure about getting __class__ in a static method) let's leave as is

[dmjoy]

Overall LGTM, happy to merge as is if you're ready.

[PaulHax]

OK, good to go by me, one less thing off my plate...