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. *
