Delta based relevance for comp reg alignment to multi-attribute targets

[jadie1]

Enables comp. reg. alignment to multi-attribute targets by adding a component step that for each probe:

1. Finds the attribute with smaller predicted delta and removes it from predicted scores
2. Alters the alignment target to only include the more relevant (bigger delta) attribute

Example call to run zeroshot:

run_align_system +experiment=phase2_june_collab/multi_attribute_pipeline_zeroshot_comparative_regression +alignment_target=june2025/ADEPT-June2025-affiliation_merit-0.0_0.0.yaml

Example call to run fewshot:

run_align_system +experiment=phase2_june_collab/multi_attribute_pipeline_fewshot_comparative_regression_loo +alignment_target=june2025/ADEPT-June2025-affiliation_merit-0.0_0.0.yaml

[eveenhuis]

Overall LGTM, I'd be fine with merging as is. Just a couple of nits/sanity checks

[eveenhuis]

Nit: David recommended I raise NotImplementedError for a similar check

[eveenhuis]

I assume we're looping all choice predictions in case different choices have different predictions? Do we think that'll actually happen?

[eveenhuis]

I know at one point we potentially had to handle both a single prediction and a list of predictions. Do we still need to do that? (If not yayyyyy simpler code :))

[eveenhuis]

This will only keep the first one if there's a tie. I think that's fine because I don't know what we'd actually do if there was a tie (plus that seems unlikely), but just double checking you didn't have a different set of assumptions

[eveenhuis]

(nit)

Suggested change

	# Update predicted scores to only have more relevant attribute
	filtered_attribute_prediction_scores = {choiceA:{}, choiceB:{}}
	for keep_attr in keep_attributes:
	filtered_attribute_prediction_scores[choiceA][keep_attr] = attribute_prediction_scores[choiceA][keep_attr]
	filtered_attribute_prediction_scores[choiceB][keep_attr] = attribute_prediction_scores[choiceB][keep_attr]
	# Update predicted scores to only have more relevant attribute
	filtered_attribute_prediction_scores = {
	choice: {
	keep_attr: predictions[choice][keep_attr]
	for keep_attr in keep_attributes
	}
	for choice in attribute_prediction_scores.keys()
	}