feat(ara): improve steering term

src/heretic/model.py

@@ -32,7 +32,7 @@ from transformers.generation import (
 )
 
 from .config import QuantizationMethod, RowNormalization, Settings
-from .utils import Prompt, batchify, empty_cache, print
+from .utils import Prompt, batchify, empty_cache, mean_distances_to_knn, print
 
 
 def get_model_class(
@@ -551,9 +551,7 @@ class Model:
                 for module_index, module in enumerate(modules):
                     # See above for a (partial) justification of this cast.
                     module = cast(Linear, module)
-
                     matrix = module.weight
-                    original_matrix = matrix.detach().clone()
 
                     good_input, good_output = good_module_io[layer_index][component][
                         module_index
@@ -563,30 +561,39 @@ class Model:
                     ]
 
                     def objective(matrix: Tensor) -> Tensor:
-                        # The results of applying the operator to inputs associated
+                        new_good_output = good_input @ matrix.T
-                        # with "good" prompts should change as little as possible.
+                        new_bad_output = bad_input @ matrix.T
+
+                        # The outputs for "good" prompts should change as little as possible.
                         preserve_good_behavior = (
-                            (good_input @ matrix.T - good_output) ** 2
+                            (new_good_output - good_output) ** 2
                         ).mean()
 
-                        # On average, the outputs for "bad" prompts should resemble
+                        # TODO: Justify the magic weights here and make them configurable.
-                        # the original outputs for "good" prompts (which steers the
+                        #       Experimentally, steer_bad_behavior needs to be about an order
-                        # behavior for "bad" prompts towards that for "good" prompts).
+                        #       of magnitude larger than preserve_good_behavior for good results.
-                        #
-                        # TODO: An alternative formulation could use the mean distance
-                        #       of "bad" outputs from the boundary of the core cluster
-                        #       of original "good" outputs. This would classify an output
-                        #       configuration as optimal as long as all "bad" outputs
-                        #       are inside the same cluster as the "good" outputs, even
-                        #       if their centroid is different from those of the "good"
-                        #       outputs.
                         steer_bad_behavior = (
-                            (
+                            0.001
-                                (bad_input @ matrix.T).mean(dim=0)
+                            # Pull the outputs for "bad" prompts towards
-                                - good_output.mean(dim=0)
+                            # the original outputs for "good" prompts.
-                            )
+                            * mean_distances_to_knn(
-                            ** 2
+                                new_bad_output,
-                        ).mean()
+                                good_output,
+                                10,
+                            ).mean()
+                            + 0.0001
+                            # Push the outputs for "bad" prompts away from
+                            # the original outputs for "bad" prompts.
+                            # In combination with the above, this overcorrects
+                            # away from the original residuals, which results
+                            # in stronger steering that can overcome more complex
+                            # refusal mechanisms.
+                            * -mean_distances_to_knn(
+                                new_bad_output,
+                                bad_output,
+                                10,
+                            ).mean()
+                        )
 
                         return (
                             preserve_good_behavior_weight * preserve_good_behavior

src/heretic/utils.py

@@ -25,6 +25,7 @@ from optuna import Trial
 from psutil import Process
 from questionary import Choice, Style
 from rich.console import Console
+from torch import Tensor
 
 from .config import DatasetSpecification, Settings
 
@@ -228,6 +229,14 @@ def batchify(items: list[T], batch_size: int) -> list[list[T]]:
     return [items[i : i + batch_size] for i in range(0, len(items), batch_size)]
 
 
+# For each vector in the 2D-tensor `a`, computes the mean Euclidean distance
+# to the `k` nearest neighbors of the vector among the vectors in the 2D-tensor `b`.
+def mean_distances_to_knn(a: Tensor, b: Tensor, k: int) -> Tensor:
+    distances = torch.cdist(a, b)
+    nearest_distances, _ = distances.topk(k, dim=1, largest=False)
+    return nearest_distances.mean(1)
+
+
 def empty_cache():
     # Collecting garbage is not an idempotent operation, and to avoid OOM errors,
     # gc.collect() has to be called both before and after emptying the backend cache.