CoolFace
Modelpublic

densenet/Gemma-4-31B-StyleTune-heretic-ara

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face
2likes89downloads
Model Card

This is a decensored version of Gryphe/Gemma-4-31B-StyleTune, made using Heretic v1.2.0 with the Arbitrary-Rank Ablation (ARA) method (with row-norm preservation)

Abliteration parameters

ParameterValue
start_layer_index30
end_layer_index48
preserve_good_behavior_weight0.8437
steer_bad_behavior_weight0.0025
overcorrect_relative_weight0.9644
neighbor_count15

Performance

MetricThis modelOriginal model ([Gryphe/Gemma-4-31B-StyleTune](https://huggingface.co/Gryphe/Gemma-4-31B-StyleTune))
KL divergence0.07330 (by definition)
Refusals8/10099/100

Uploaded finetuned model

  • —Developed by: densenet
  • —License: apache-2.0
  • —Finetuned from model : Gryphe/Gemma-4-31B-StyleTune

This gemma4 model was trained 2x faster with Unsloth and Huggingface's TRL library.

<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>