Downloads · 30 days
34
26% of all-time downloads
Gemma4-E4B-8B-Heretic-Ultra
dopaemon/Gemma4-E4B-8B-Heretic-Ultra
Gemma4-E4B-8B-Heretic-Ultra is a any-to-any model from dopaemon. Use it for the any-to-any task on the model card, and read the license before you ship it in a product. It is set up for transformers. The card lists the license as apache-2.0.
A decensored version of google/gemma-4-E4B-it, abliterated using Heretic v1.2.0 with the Arbitrary-Rank Ablation (ARA) method.
All-time downloads
131
Public
Parameters
8B
16 GB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors16 GB · 100%
At a glance
- Task
- Any-to-Any
- Library
- transformers
- License
- apache-2.0
- Model type
- gemma4
- Access
- Public
- Created
- Sep 23, 2026
- Updated
- Sep 23, 2026
- SHA
- 81813ded
Base models
Model card
From the Hugging Face model README
Gemma4 E4B 8B Heretic Ultra
A decensored version of google/gemma-4-E4B-it, abliterated using Heretic v1.2.0 with the Arbitrary-Rank Ablation (ARA) method.
Credits
- Base model: google/gemma-4-E4B-it by Google DeepMind.
- Abliteration / decensoring: this specific Heretic Ultra abliteration was created by llmfan46 using the Heretic tool. All credit for the decensoring work goes to them.
- This repo: hosts the original fp16 weights (
main) and an MLX 4-bit quantized conversion (MLX-4Bitbranch) for use with mlx-lm on Apple Silicon.
Abliteration parameters
| Parameter | Value |
|---|---|
| start_layer_index | 7 |
| end_layer_index | 36 |
| preserve_good_behavior_weight | 0.5783 |
| steer_bad_behavior_weight | 0.0001 |
| overcorrect_relative_weight | 0.9986 |
| neighbor_count | 15 |
Targeted components
- attn.o_proj
Performance
| Metric | This model | Original model (gemma-4-E4B-it) |
|---|---|---|
| KL divergence | 0.0076 | 0 (by definition) |
| Refusals | 3/100 | 99/100 |
PIQA
| Original | Heretic | |
|---|---|---|
| Total questions | 1838 | 1838 |
| Correct | 1581 | 1578 |
| Accuracy | 86.02% | 85.85% |
MMLU
| Original | Heretic | |
|---|---|---|
| Total questions | 14042 | 14042 |
| Correct | 9753 | 9633 |
| Accuracy | 69.46% | 68.60% |
Branches
| Branch | Contents |
|---|---|
main | Original fp16 weights (safetensors) |
MLX-4Bit | MLX 4-bit quantized version, converted for mlx-lm |
Usage
from transformers import AutoProcessor, AutoModelForCausalLM
MODEL_ID = "dopaemon/Gemma4-E4B-8B-Heretic-Ultra"
processor = AutoProcessor.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(
MODEL_ID,
dtype="auto",
device_map="auto"
)
See google/gemma-4-E4B-it for the full base model card, architecture details, and license terms.
- Task
- Any-to-Any
- Library
- transformers
- Type
- gemma4
- License
- apache-2.0
- Created
- Sep 23, 2026
- Updated
- Sep 23, 2026