Ty_Tech/heretic
4
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Switch to multi-objective optimization

Browse Source
This commit is contained in:
Philipp Emanuel Weidmann
2025-11-14 18:04:23 +05:30
parent 0bae27f359
commit 8a1aceff11
6 changed files with 214 additions and 203 deletions
Download Patch File Download Diff File Expand all files Collapse all files
config.default.toml
+3 -6
View File
@@ -24,12 +24,9 @@ max_batch_size = 128
# Maximum number of tokens to generate for each response.
max_response_length = 100
# Maximum Kullback-Leibler divergence from the original model to allow for abliterated models.
max_kl_divergence = 0.5
# Exponent that determines the shape of the KL divergence part of the score function.
# See evaluator.py for the exact meaning of this parameter.
kl_score_shape = 3.0
# Assumed "typical" value of the Kullback-Leibler divergence from the original model for abliterated models.
# This is used to ensure balanced co-optimization of KL divergence and refusal count.
kl_divergence_scale = 1.0
# Number of abliteration trials to run during optimization.
n_trials = 200