Fully automatic censorship removal for language models
Support multimodal models
Philipp Emanuel Weidmann8b827ee386ac
diff --git a/src/heretic/main.py b/src/heretic/main.pyindex 687fec0..ec8f41c 100644--- a/src/heretic/main.py+++ b/src/heretic/main.py@@ -177,11 +177,11 @@ def run():max_weight = trial.suggest_float("max_weight", 0, 1)max_weight_position = trial.suggest_float(- "max_weight_position", 0, len(model.model.model.layers) - 1+ "max_weight_position", 0, len(model.get_layers()) - 1)min_weight = trial.suggest_float("min_weight", 0, max_weight)min_weight_distance = trial.suggest_float(- "min_weight_distance", 1, len(model.model.model.layers) - 1+ "min_weight_distance", 1, len(model.get_layers()) - 1)print()@@ -226,10 +226,10 @@ def run():{"max_weight": max_weight,"max_weight_position": max_weight_position- * (len(model.model.model.layers) - 1),+ * (len(model.get_layers()) - 1),"min_weight": min_weight,"min_weight_distance": min_weight_distance- * (len(model.model.model.layers) - 1),+ * (len(model.get_layers()) - 1),})diff --git a/src/heretic/model.py b/src/heretic/model.pyindex 0d30f47..0a618c9 100644--- a/src/heretic/model.py+++ b/src/heretic/model.py@@ -7,6 +7,7 @@ from typing import Anyimport torchimport torch.nn.functional as Ffrom torch import LongTensor+from torch.nn import ModuleListfrom transformers import (AutoModelForCausalLM,AutoTokenizer,@@ -59,9 +60,7 @@ class Model:if self.model is None:raise Exception("Failed to load model with all configured dtypes.")- print(- f"* Transformer model with [bold]{len(self.model.model.layers)}[/] layers"- )+ print(f"* Transformer model with [bold]{len(self.get_layers())}[/] layers")print(f"* [bold]{len(self.get_layer_matrices(0))}[/] abliterable matrices per layer")@@ -79,8 +78,16 @@ class Model:device_map=self.settings.device_map,)+ def get_layers(self) -> ModuleList:+ # Most multimodal models.+ with suppress(Exception):+ return self.model.model.language_model.layers++ # Text-only models.+ return self.model.model.layers+def get_layer_matrices(self, layer_index: int) -> list[torch.Tensor]:- layer = self.model.model.layers[layer_index]+ layer = self.get_layers()[layer_index]matrices = []@@ -118,7 +125,7 @@ class Model:):# Note that some implementations of abliteration also orthogonalize# the embedding matrix, but it's unclear if that has any benefits.- for layer_index in range(len(self.model.model.layers)):+ for layer_index in range(len(self.get_layers())):distance = abs(layer_index - max_weight_position)# Don't orthogonalize layers that are more than