Merge branch 'master' into e2e-tests

src/heretic/main.py

@@ -5,6 +5,14 @@
 
 import sys
 
+# Ensure standard output/error use UTF-8 instead of system default charmap (e.g. cp1252 on Windows).
+for stream in (sys.stdout, sys.stderr):
+    if (
+        hasattr(stream, "reconfigure")
+        and (getattr(stream, "encoding", "") or "").lower() != "utf-8"
+    ):
+        stream.reconfigure(encoding="utf-8")  # type: ignore
+
 from .config import Settings
 
 
@@ -590,10 +598,22 @@ def run():
             # The parameter ranges are based on experiments with various models
             # and much wider ranges. They are not set in stone and might have to be
             # adjusted for future models.
-            max_weight = trial.suggest_float(
+            #
-                f"{component}.max_weight",
+            # The MLP gets a negative lower bound that is then clamped to 0, so the
-                0.8,
+            # optimizer can fully disable its ablation. The clamp puts a positive
-                1.5,
+            # probability mass on exactly 0 (the continuous sampler would otherwise
+            # reach 0 with probability zero). Ablating the MLP is often unnecessary for
+            # removing refusals and tends to damage model intelligence more than
+            # ablating the attention output, so on many models the optimum is to leave
+            # it (mostly) untouched. See issue #202.
+            max_weight_lower_bound = -0.25 if component == "mlp.down_proj" else 0.8
+            max_weight = max(
+                0.0,
+                trial.suggest_float(
+                    f"{component}.max_weight",
+                    max_weight_lower_bound,
+                    1.5,
+                ),
             )
             max_weight_position = trial.suggest_float(
                 f"{component}.max_weight_position",

src/heretic/model.py

@@ -499,6 +499,12 @@ class Model:
                     params.min_weight - params.max_weight
                 )
 
+                # A weight of 0 disables this component's ablation. reset_model() has
+                # already left the adapter at identity, so abort before the otherwise
+                # wasteful decomposition (which would also be operating on a zero matrix).
+                if weight == 0:
+                    continue
+
                 if refusal_direction is None:
                     # The index must be shifted by 1 because the first element
                     # of refusal_directions is the direction for the embeddings.