NullpoLab's picture
Upload folder using huggingface_hub
23bdb7c verified
|
Raw
History Blame
2.51 kB
metadata
library_name: transformers
license: apache-2.0
license_link: https://ai.google.dev/gemma/docs/gemma_4_license
pipeline_tag: image-text-to-text
base_model:
  - google/gemma-4-E4B-it
tags:
  - heretic
  - uncensored
  - decensored
  - abliterated
  - ara

gemma-4-E4B-it-Heretic-ARA-Refusals8

Overview

This is a decensored version of google/gemma-4-E4B-it, made using Heretic v1.2.0 with the Arbitrary-Rank Ablation (ARA) method.

Abliteration Method

ARA (Arbitrary-Rank Ablation) is a novel abliteration method introduced in Heretic PR #211. Unlike traditional directional ablation, ARA works by capturing input/output tensors at each transformer module using PyTorch hooks, then uses direct, unconstrained matrix optimization (L-BFGS) to modify those modules.

The optimization balances three competing goals:

  1. Preserve outputs for "harmless" prompts as much as possible
  2. Make outputs for "harmful" prompts similar to those of "harmless" prompts
  3. Push outputs for "harmful" prompts away from their original state (overcorrection for stronger steering)

Abliteration Parameters

Parameter Value
start_layer_index 21
end_layer_index 35
preserve_good_behavior_weight 0.9896
steer_bad_behavior_weight 0.0010
overcorrect_relative_weight 0.4414
neighbor_count 15

Performance

Metric This model Original model (google/gemma-4-E4B-it)
Refusals 8/100 99/100
KL divergence 0.0216 0 (by definition)

Evaluation was conducted using mlabonne/harmful_behaviors (100 test prompts) for refusal count, and mlabonne/harmless_alpaca (100 test prompts) for KL divergence.

Notes

  • This model is intended for research and creative writing purposes
  • The ARA branch of Heretic (PR #211) is currently in Draft status and not yet merged into the main branch (as of April 4, 2026; commit: 3b70fe5)
  • Refusal suppression was evaluated using English prompts only; behavior on Japanese prompts may differ
  • Base model: google/gemma-4-E4B-it
  • Heretic: p-e-w/heretic