Fix sign toggle in removal-based interpretability metrics

[DarylOkeke]

RemovalBasedMetric.compute() aliased y_probs instead of copying it, so the in-place negation of NEGATIVE-class samples re-flipped the sign on every percentage in the loop — the score at 20% comes out 0.0113 with percentages=[20] but 1.0088 with [10, 20]

clone y_probs and ablated_probs before negating (the negation itself is correct, it just needed a copy), plus a test that a percentage's score doesn't depend on its position in the list. Fails at 0.997 absolute difference without the fix

only affects filters that emit SampleClass.NEGATIVE, so the default threshold_sample_filter is fine but examples/interpretability/custom_sample_filter.py isn't

note: could also hoist the negation out of the loop since it doesn't depend on the percentage, kept it to the two clones to stay small

[lehendo]

Not approved actually. Pls look at the comments below.

[lehendo]

Actually wait, can you make an extra change one sec.

Can you just add this to your PR:

neg_mask = evaluated_drops < 0 inside the if debug and return_per_percentage: block (~line 496) reuses the neg_mask name from line 402, which is read again on the next loop iteration (lines 457/460).

since if/for don't create new scopes, debug=True with multiple percentages would silently clobber the real mask mid-loop.

[lehendo]

Also unsure why your PR isn't triggering the PR checks. Make sure your branch is updated and all of that.

[DarylOkeke]

Actually wait, can you make an extra change one sec.

Can you just add this to your PR:

neg_mask = evaluated_drops < 0 inside the if debug and return_per_percentage: block (~line 496) reuses the neg_mask name from line 402, which is read again on the next loop iteration (lines 457/460).

since if/for don't create new scopes, debug=True with multiple percentages would silently clobber the real mask mid-loop.

Fixed.