---
license: apache-2.0
base_model: google/gemma-4-E4B-it
tags: [heretic, abliterated, uncensored, gemma4]
pipeline_tag: any-to-any
---

# Gemma-4-1-E4B-it (heretic)

Abliterated (refusal-direction ablated) build of the **refreshed** Gemma-4 E4B-it.

"**Gemma-4-1**" is our label for the **2026-07-20 refreshed** Gemma-4 weights
(post chat-template / tool-calling fixes) — it is **not** an official Google
release name. The true base is `google/gemma-4-E4B-it` at commit `ee0ef6023621cff504d758262d4e04895a5af4a2`.

Decensored with [heretic](https://github.com/p-e-w/heretic) (optimized
abliteration; dual objective: minimise refusals + minimise KL from the base).
Base license Apache 2.0. No fine-tuning; weights-only intervention.

## This build (Pareto trial 198)
- Refusals **25/100**, KL divergence **0.0219** on heretic's harmful_behaviors eval —
  the low-KL knee of the 200-trial front (near-max refusal reduction, minimal capability drift).

## Reproduce any trial
`heretic-study.jsonl` in this repo is the complete 200-trial Optuna study. To
publish a more-aggressive point (e.g. trial 106: 23/100 refusals, KL 0.133):
```
mkdir checkpoints && cp heretic-study.jsonl checkpoints/google--gemma-4-E4B-it.jsonl
heretic --model google/gemma-4-E4B-it --study-checkpoint-dir checkpoints
```
This reopens the Pareto menu with zero re-search.