Clone of https://huggingface.co/Sabomako/gemma-4-31B-it-heretic made for quantization purposes with unique name
This is a decensored version of google/gemma-4-31B-it, made using Heretic v1.3.0 with custom ORBA abliteration method
Refusal classification via an LLM judge (Gemma-12B-IT-Heretic).
Performance
| Metric | This model | Original model (Qwen/Qwen3.6-27B) |
|---|---|---|
| KL divergence | 0.08095 | 0 (by definition) |
| Refusals | 2/100 | 98/100 |
- Downloads last month
- 28