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 (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.

Downloads last month
190
Safetensors
Model size
8B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for eshucode/Gemma-4-1-E4B-it-heretic

Finetuned
(349)
this model
Quantizations
3 models