Suggestion: start-end layer probes before start

[kabachuha]

Currently, in the processes such as ARA Heretic optimizes for six parameters

Parameters:

• start_layer_index = 11
• end_layer_index = 34
• preserve_good_behavior_weight = 0.5510
• steer_bad_behavior_weight = 0.0210
• overcorrect_relative_weight = 0.9777
• neighbor_count = 13

While the four latest are mathematical and ARA related, the first two seem to be offtopic and they are completely independent from ARA, potentially hindering its optimization landscape.

I wonder if Heretic can firstly perform a brief optimal start end layers fit check and only then proceed to refine the math coefficients of ARA.

[p-e-w]

How would that work? The layer range is NOT independent from the other parameters because we optimize for the overall model behavior. All parameters are interrelated.

We could try heuristics like discarding layers that have low silhouette scores on the residual vectors, indicating that the layer doesn't clearly distinguish refusals from non-refusals, but that is at best a rough first guess.