diff --git a/src/heretic/model.py b/src/heretic/model.py index f3b3c09..9feaab9 100644 --- a/src/heretic/model.py +++ b/src/heretic/model.py @@ -576,11 +576,11 @@ class Model: (new_good_output - good_output) ** 2 ).mean() - # TODO: Justify the magic weights here and make them configurable. + # TODO: Justify the magic weights here and make them configurable/optimizable. # Experimentally, steer_bad_behavior needs to be about an order # of magnitude larger than preserve_good_behavior for good results. steer_bad_behavior = ( - 0.001 + 1.0 # Pull the outputs for "bad" prompts towards # the original outputs for "good" prompts. * mean_distances_to_knn( @@ -588,7 +588,7 @@ class Model: good_output, 10, ).mean() - + 0.0001 + + 0.5 # Push the outputs for "bad" prompts away from # the original outputs for "bad" prompts. # In combination with the above, this overcorrects