Downloads · 30 days
19
16% of all-time downloads
gemma-3-27b-it-heretic-v1.2
grayarea/gemma-3-27b-it-heretic-v1.2
gemma-3-27b-it-heretic-v1.2 is a machine learning model from grayarea. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
This is a decensored version of gemma-3-27b-it, made using Heretic v1.2.0 focusing on zero refusals with low KL divergence
All-time downloads
118
Public
Parameters
27.4B
54.9 GB on disk
Likes
3
Public
Repo makeup
Click a slice to open those files.
.safetensors54.9 GB · 100%
At a glance
- Model type
- gemma3
- Access
- Public
- Created
- Mar 8, 2026
- Updated
- Mar 8, 2026
- SHA
- 69cf90a7
Base models
Model card
From the Hugging Face model README
This is a decensored version of gemma-3-27b-it, made using Heretic v1.2.0 focusing on zero refusals with low KL divergence
KL Divergence
| Metric | This Model | Original Model |
|---|---|---|
| KL divergence | 0.0539 | 0 (by definition) |
| Refusals | 0/108 | 107/108 |
Abliteration parameters
- Zero refusals with KL divergence of 0.0539
- Custom heretic training dataset
- Model targetted heretic configuration
- Abliterated with MPOA enabled (Magnitude-Preserving Orthogonal Ablation)
- Full row renormalization
- Winsorization Quantile 0.997
The following benchmarks are for the quantized version of this model
Relative Perplexity
| Quant | Filename | PPL ± Error |
|---|---|---|
| Q8_0 | gemma-3-27b-it-Q8_0.gguf (original baseline) | 6.5517 +/- 0.04674 |
| Q8_0 | gemma-3-27b-it-heretic-v1.2-Q8_0.gguf | 6.5510 +/- 0.04666 |
| Q4_K_M | gemma-3-27b-it-heretic-v1.2-Q4_K_M.gguf | 6.6245 +/- 0.04704 |
Benchmark Comparison
| Benchmark | gemma-3-27b-it-Q8_0.gguf | gemma-3-27b-it-Q4_K_M.gguf | gemma-3-27b-it-heretic-v1.2-Q4_K_M.gguf |
|---|---|---|---|
| Perplexity (Wikitext-2) | 6.5517 | 6.6226 | 6.6245 |
| HellaSwag | 81.75% | 81.75% | 80.50% |
| Winogrande | 76.24% | 76.32% | 76.40% |
| ARC-Challenge | 57.86% | 58.86% | 59.53% |
| MMLU | 45.35% | 46.58% | 45.41% |
*Note: MMLU benchmark has moral_scenarios, moral_disputes, business_ethics, professional_law and jurisprudence subjects removed. *
- Type
- gemma3
- Created
- Mar 8, 2026
- Updated
- Mar 8, 2026