license: apache-2.0
language:
- zh
- en
pipeline_tag: text-generation
tags: - heretic
- uncensored
- ablation
- qwen
base_model: - Qwen/Qwen3.5-27B
Evaluation & Comparison
Core Metrics
| Metric | Original Model | Heretic Model | Description |
|---|---|---|---|
| Refusal Rate | 92.0% | 2/100 | Tested on 520 harmful prompts |
| KL Divergence | - | 0.0414 | Sequence cumulative KL (per token) |
| NLL Change | - | +4.2% | Minor impact on language capability |
| Model Size | 27B | 27B | Architecture unchanged |
KL Divergence Rating
KL divergence measures the degree of model modification:
| KL Range | Rating | Description |
|---|---|---|
| < 0.05 | ⭐⭐⭐⭐⭐ | Extremely Low - Model virtually unchanged |
| 0.05 - 0.10 | ⭐⭐⭐⭐ | Low - Minor modification, capabilities well preserved |
| 0.10 - 0.20 | ⭐⭐⭐ | Moderate - Acceptable modification range |
| 0.20 - 0.50 | ⭐⭐ | High - Possible noticeable capability loss |
| > 0.50 | ⭐ | Too High - Model may be severely compromised |
This model: KL = 0.0414 , Refusal Rate: 2/100 , NLL : +4.2%
Residual Visualization
PaCMAP projections showing the mixing of harmless (blue) and harmful (red) prompts:
| Layer 20 | Layer 30 |
|---|---|
 |
 |
| Layer 40 | Layer 55 |
|---|---|
 |
 |
These plots show successful removal of refusal behavior - harmless and harmful prompts are well-mixed across layers.
Technical Method
ABLIteration Approach
This model uses the Heretic ABLIteration method for neural direction ablation:
- Identify Refusal Direction - Train a LoRA on harmful behavior datasets to identify neural directions controlling "refusal behavior"
- Direction Extraction - Extract the "refusal vector" from the trained LoRA
- Ablative Removal - Subtract this direction from the original model weights, removing the censorship mechanism
This method only modifies model weights without changing the architecture or adding inference overhead.
For detailed technical principles, refer to: Heretic Abliteration
Data Sources
| Purpose | Dataset |
|---|---|
| Refusal Direction Identification | mlabonne/harmful_behaviors (520 prompts) |
| KL Evaluation | General prompts (100 prompts) |
| Refusal Rate Testing | mlabonne/harmful_behaviors (520 prompts) |
✅ Recommended Uses
- Research and analysis of sensitive topics
- Safety testing and red-teaming exercises
- Academic research on model alignment
❌ Not Recommended For
- Production environments requiring content moderation
- Applications targeting minors
- Scenarios with potential legal risks
Limitations
- Minor Capability Loss - NLL increased by approximately 4.2%, which may slightly affect performance on complex tasks
- User Discretion Required - Users must independently judge the appropriateness of generated outputs
Disclaimer
⚠️ Important: This model is intended for research and educational purposes only.
- This model has had its censorship mechanisms removed and may generate harmful, dangerous, or inappropriate content
- Users assume all risks associated with usage
- Do not use this model for illegal activities, harming others, or any inappropriate purposes
- The model authors are not liable for any indirect, incidental, or consequential damages
Acknowledgments
- Original Model: Qwen/Qwen3.5-27B
- Heretic Method: alignlab/heretic