Fully automatic censorship removal for language models
Support Phi-3.5-MoE
Philipp Emanuel Weidmann7ed0cb1ffba4
diff --git a/src/heretic/model.py b/src/heretic/model.pyindex 0a618c9..ae24837 100644--- a/src/heretic/model.py+++ b/src/heretic/model.py@@ -106,6 +106,12 @@ class Model:for expert in layer.mlp.experts:try_add(expert.down_proj.weight)+ # Phi-3.5-MoE (and possibly others).+ if not matrices:+ with suppress(Exception):+ for expert in layer.block_sparse_moe.experts:+ try_add(expert.w2.weight)+# We need at least one MLP down-projection.assert matrices