feat: make adjective describing the type of model produced configurable (#539)

config.default.toml

@@ -87,7 +87,7 @@ modifiers = [
     { plugin = "heretic.modifiers.ara.ARA" },
 ]
 
-# Number of abliteration trials to run during optimization.
+# Number of trials to run during optimization.
 n_trials = 100
 
 # Number of trials that use random sampling for the purpose of exploration.
@@ -96,6 +96,9 @@ n_startup_trials = 30
 # Directory to save and load study progress to/from.
 study_checkpoint_dir = "checkpoints"
 
+# Adjective describing the type of model produced by the configuration.
+modified_model_type = "decensored"
+
 # Maximum size for individual safetensors files generated when exporting a model.
 max_shard_size = "5GB"
 

config.nohumor.toml

@@ -3,6 +3,8 @@
 
 max_response_length = 300
 
+modified_model_type = "humor-reduced"
+
 upload_tags = [
     "no-humor",
 ]

config.noslop.toml

@@ -3,6 +3,8 @@
 
 max_response_length = 300
 
+modified_model_type = "slop-reduced"
+
 upload_tags = [
     "no-slop",
 ]

config.nosyco.toml

@@ -3,6 +3,8 @@
 
 max_response_length = 100
 
+modified_model_type = "sycophancy-reduced"
+
 upload_tags = [
     "no-sycophancy",
 ]

src/heretic/config.py

@@ -201,7 +201,7 @@ class Settings(BaseSettings):
     evaluate_model: str | None = Field(
         default=None,
         description=(
-            "If this model ID or path is set, then instead of abliterating the main model, "
+            "If this model ID or path is set, then instead of modifying the main model, "
             "evaluate this model relative to the main model."
         ),
         exclude=True,
@@ -210,7 +210,7 @@ class Settings(BaseSettings):
     collect_reproducibles: str | None = Field(
         default=None,
         description=(
-            "If this directory path is set, then instead of abliterating a model, "
+            "If this directory path is set, then instead of modifying a model, "
             "download all reproduce.json files from public Heretic model repositories "
             "on Hugging Face, and store them in that directory for archival purposes."
         ),
@@ -221,7 +221,7 @@ class Settings(BaseSettings):
         default=None,
         description=(
             "If this path or URL to a reproduce.json file is set, load reproduction information "
-            "from that file, and attempt to reproduce the abliterated model it originated from."
+            "from that file, and attempt to reproduce the modified model it originated from."
         ),
         exclude=True,
     )
@@ -405,7 +405,7 @@ class Settings(BaseSettings):
 
     n_trials: PositiveInt = Field(
         default=100,
-        description="Number of abliteration trials to run during optimization.",
+        description="Number of trials to run during optimization.",
     )
 
     n_startup_trials: NonNegativeInt = Field(
@@ -427,6 +427,11 @@ class Settings(BaseSettings):
         exclude=True,
     )
 
+    modified_model_type: str = Field(
+        default="decensored",
+        description="Adjective describing the type of model produced by the configuration.",
+    )
+
     benchmarks: list[BenchmarkSpecification] = Field(
         default=[
             BenchmarkSpecification(
@@ -485,7 +490,7 @@ class Settings(BaseSettings):
                 description="An Adversarial Winograd Schema Challenge at Scale",
             ),
         ],
-        description="Benchmarks to offer to the user for evaluating abliterated models.",
+        description="Benchmarks to offer to the user for evaluating modified models.",
         exclude=True,
     )
 
@@ -516,7 +521,7 @@ class Settings(BaseSettings):
 
     model_action: str | None = Field(
         default=None,
-        description='Action to take with the decensored model: "save", "upload", or unset to prompt the user.',
+        description='Action to take with the modified model: "save", "upload", or unset to prompt the user.',
     )
 
     save_directory: str | None = Field(

src/heretic/main.py

@@ -175,7 +175,7 @@ def obtain_export_strategy(
             "How do you want to export the model?",
             choices=[
                 Choice(
-                    title="Merge the abliteration LoRA and export the full model"
+                    title="Merge the LoRA and export the full model"
                     + (
                         ""
                         if settings.quantization == QuantizationMethod.NONE
@@ -184,7 +184,7 @@ def obtain_export_strategy(
                     value=ExportStrategy.MERGE,
                 ),
                 Choice(
-                    title="Export the abliteration LoRA only (can be merged later)",
+                    title="Export the LoRA only (can be merged later)",
                     value=ExportStrategy.ADAPTER,
                 ),
             ],
@@ -944,7 +944,7 @@ def run():
                 action = ask_if_unset(
                     settings.model_action,
                     questionary.select(
-                        "What do you want to do with the decensored model?",
+                        f"What do you want to do with the {settings.modified_model_type} model?",
                         choices=[
                             Choice(
                                 title="Save the model to a local folder",
@@ -1366,11 +1366,11 @@ def run():
 
                             scope = questionary.select(
                                 (
-                                    "Do you want to benchmark the original model along with the decensored model? "
+                                    f"Do you want to benchmark the original model along with the {settings.modified_model_type} model? "
                                     "Benchmarking both models allows you to compare the scores, but it takes twice as much time."
                                 ),
                                 choices=[
-                                    "Benchmark only the decensored model",
+                                    f"Benchmark only the {settings.modified_model_type} model",
                                     "Benchmark both models",
                                 ],
                                 style=Style([("highlighted", "reverse")]),

src/heretic/model.py

@@ -242,7 +242,7 @@ class Model:
                     all_components[component] = 0
                 all_components[component] += len(modules)
 
-        print("* Abliterable components:")
+        print("* Modifiable components:")
         for component, count in all_components.items():
             print(f"  * [bold]{component}[/]: [bold]{count}[/] modules total")
 
@@ -250,7 +250,7 @@ class Model:
         # Guard against calling this method at the wrong time.
         assert isinstance(self.model, PreTrainedModel)
 
-        # Always use LoRA adapters for abliteration (faster reload, no weight modification).
+        # Always use LoRA adapters for modification (faster reload, no weight modification).
         # Collect actual leaf module names from the model for LoRA targeting.
         # This is more robust than splitting component keys (e.g. "attn.o_proj" -> "o_proj")
         # because hybrid models like Qwen3.5 MoE have modules with different names
@@ -513,13 +513,13 @@ class Model:
             for expert in layer.moe.experts:  # ty:ignore[not-iterable, unresolved-attribute]
                 try_add("mlp.down_proj", expert.output_linear)  # ty: ignore[unresolved-attribute]
 
-        # We need at least one module across all components for abliteration to work.
+        # We need at least one module across all components.
         total_modules = sum(len(mods) for mods in modules.values())
-        assert total_modules > 0, "No abliterable modules found in layer"
+        assert total_modules > 0, "No modifiable modules found in layer"
 
         return modules
 
-    def get_abliterable_components(self) -> list[str]:
+    def get_modifiable_components(self) -> list[str]:
         components: set[str] = set()
 
         # Scan all layers because hybrid models (e.g. Qwen3.5 MoE) have different

src/heretic/modifiers/abliteration.py

@@ -255,7 +255,7 @@ class Abliteration(Modifier[Parameters]):
 
         weight_distributions = {}
 
-        for component in model.get_abliterable_components():
+        for component in model.get_modifiable_components():
             # 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.

src/heretic/utils.py

@@ -344,7 +344,7 @@ def get_readme_intro(
     else:
         reproducibility_instructions = ""
 
-    return f"""# This is a decensored version of {
+    return f"""# This is a {settings.modified_model_type} version of {
         model_link
     }, made using [Heretic](https://heretic-project.org) v{version("heretic-llm")}
 {reproducibility_instructions}