Uncensored Gemma 4 MLX
Collection
10 items • Updated • 2
Gemma-4-26B-A4B Heretic
Quality: quantized (mixed quants per tensor, group size: 32, 9.367 bpw)
Most layers use 8-bit affine quantization with a group size 32; some layers are saved in bf16.
This is a abliterated (uncensored) version of google/gemma-4-26B-A4B-it, made using Heretic v1.2.0 with the Arbitrary-Rank Ablation (ARA) method (with row-norm preservation)
Performance
| Metric | This model | Original model (google/gemma-4-26B-A4B-it) |
|---|---|---|
| KL divergence | 0.0499 | 0 (by definition) |
| Refusals | 11/100 | 100/100 |
Abliteration parameters
| Parameter | Value |
|---|---|
| start_layer_index | 10 |
| end_layer_index | 30 |
| preserve_good_behavior_weight | 0.5480 |
| steer_bad_behavior_weight | 0.0009 |
| overcorrect_relative_weight | 0.5868 |
| neighbor_count | 14 |
Source
This model was converted to MLX format from coder3101/gemma-4-26B-A4B-it-heretic using mlx-vlm version 0.4.4.
- Downloads last month
- 187
Model size
27B params
Tensor type
BF16
·
U32 ·
Hardware compatibility
Log In to add your hardware
8-bit
Model tree for TheCluster/Gemma-4-26B-A4B-Heretic-MLX-9bit
Base model
google/gemma-4-26B-A4B Finetuned
google/gemma-4-26B-A4B-it Finetuned
coder3101/gemma-4-26B-A4B-it-heretic