Gemma 3 12B IT Abliterated
| Property | Value |
|---|---|
| Model Size | 12B parameters |
| Base Model | google/gemma-3-12b-it |
| Author | mlabonne |
| Model URL | HuggingFace |
| Recommended Parameters | temperature=1.0, top_k=64, top_p=0.95 |
What is gemma-3-12b-it-abliterated-GGUF?
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 maintaining model capabilities while removing traditional response restrictions, achieving an impressive >90% acceptance rate for queries while preserving coherent output generation.
Implementation Details
The model employs a novel layerwise abliteration approach, applying refusal direction computations across multiple layers (3 to 45) independently. This technique differs from traditional methods by focusing on hidden states rather than residual streams, combined with a refusal weight of 0.6 to optimize the balance between unrestricted responses and output quality.
- Implements layerwise abliteration across 42 model layers
- Uses a 0.6 refusal weight for optimal performance
- Demonstrates higher resilience to abliteration compared to models like Qwen 2.5
- Experimental implementation with occasional text artifacts
Core Capabilities
- High acceptance rate (>90%) for previously restricted queries
- Maintained coherent output generation
- Optimized for specific generation parameters
- Balanced performance between freedom and quality
Frequently Asked Questions
Q: What makes this model unique?
The model's distinctive feature is its novel layerwise abliteration technique, which achieves uncensored outputs while maintaining high coherence. Unlike traditional approaches, it processes refusal directions layer by layer, resulting in more nuanced and controlled removal of restrictions.
Q: What are the recommended use cases?
This model is best suited for applications requiring unrestricted responses while maintaining output quality. Users should be aware of occasional text artifacts and use the recommended generation parameters (temperature=1.0, top_k=64, top_p=0.95) for optimal results.
