add configured max tokens for refusal test generation. Exposes --refusal-max-tokens via CLI and the abliterator constructor (default to 128, not changed). Better for resoning models like qwen3. :>

obliteratus/abliterate.py

@@ -843,6 +843,7 @@ class AbliterationPipeline:
         max_seq_length: int | None = None,
         # Verify stage sample size
         verify_sample_size: int | None = None,
+        refusal_max_tokens: int | None = None,
         on_stage: Callable[[StageResult], None] | None = None,
         on_log: Callable[[str], None] | None = None,
     ):
@@ -972,6 +973,7 @@ class AbliterationPipeline:
         # refusal rate measurement.  Default 30 gives ~3.3% resolution;
         # increase for tighter confidence intervals (reviewer feedback).
         self.verify_sample_size = verify_sample_size if verify_sample_size is not None else 30
+        self.refusal_max_tokens = refusal_max_tokens if refusal_max_tokens is not None else 128
 
         # Large model mode: conservative defaults for 120B+ models.
         # Reduces memory footprint by limiting SAE features, directions,
@@ -6338,7 +6340,7 @@ class AbliterationPipeline:
                     with torch.no_grad():
                         outputs = model.generate(
                             **inputs,
-                            max_new_tokens=128,
+                            max_new_tokens=self.refusal_max_tokens,
                             do_sample=False,
                         )
 

obliteratus/cli.py

@@ -240,6 +240,10 @@ def main(argv: list[str] | None = None):
             help="Number of harmful prompts to test for refusal rate (default: 30). "
                  "Increase for tighter confidence intervals (e.g. 100 for ~1%% resolution).",
         )
+        p.add_argument(
+            "--refusal-max-tokens", type=int, default=None,
+            help="Max new tokens to generate per response in the refusal test (default: 128).",
+        )
         p.add_argument(
             "--dataset", type=str, default="builtin",
             help="Prompt dataset source for contrastive extraction when using residue mining (default: builtin).",
@@ -1043,6 +1047,7 @@ def _cmd_abliterate(args):
         quantization=args.quantization,
         large_model_mode=getattr(args, "large_model", False),
         verify_sample_size=getattr(args, "verify_sample_size", None),
+        refusal_max_tokens=getattr(args, "refusal_max_tokens", None),
         on_stage=on_stage,
         on_log=on_log,
         **prompt_kwargs,
@@ -1368,6 +1373,8 @@ def _cmd_remote_abliterate(args):
         kwargs["large_model"] = True
     if getattr(args, "verify_sample_size", None) is not None:
         kwargs["verify_sample_size"] = args.verify_sample_size
+    if getattr(args, "refusal_max_tokens", None) is not None:
+        kwargs["refusal_max_tokens"] = args.refusal_max_tokens
 
     result_path = runner.run_obliterate(
         model=args.model,