| Property | Value |
|---|---|
| Model Size | 12B parameters |
| Base Model | google/gemma-3-12b-it |
| Author | mlabonne |
| Model Hub | HuggingFace |
| Recommended Parameters | temperature=1.0, top_k=64, top_p=0.95 |
What is gemma-3-12b-it-abliterated?
This is an experimental uncensored version of Google's Gemma 3 12B model, created using an innovative layerwise abliteration technique. The model represents a significant advancement in reducing AI safety restrictions while maintaining core model capabilities.
Implementation Details
The model employs a novel layerwise abliteration approach, processing layers 3 through 45 independently. It utilizes a refusal direction computation based on hidden states, combined with a 0.6 refusal weight to enhance the acceptance rate beyond 90% while preserving output coherence.
- Implements layerwise abliteration across multiple model layers
- Uses hidden state-based refusal direction computation
- Applies a 0.6 refusal weight for balanced output
- Demonstrates high resilience to abliteration compared to other models
Core Capabilities
- High acceptance rate (>90%) for previously restricted content
- Maintained coherent output generation
- Experimental text processing with occasional minor artifacts
- Optimized performance with specific generation parameters
Frequently Asked Questions
Q: What makes this model unique?
This model introduces a new approach to abliteration by processing individual layers independently, resulting in a high acceptance rate while maintaining output quality. It's particularly notable for its resilience to abliteration compared to other models like Qwen 2.5.
Q: What are the recommended use cases?
The model is best suited for applications requiring reduced content restrictions while maintaining coherent outputs. Users should be aware of potential occasional text artifacts and use the recommended generation parameters for optimal results.