Gemma-3-4B-IT Abliterated
| Property | Value |
|---|---|
| Base Model | google/gemma-3-4b-it |
| Parameter Count | 4 Billion |
| Hugging Face | Model Repository |
| Author | mlabonne |
What is gemma-3-4b-it-abliterated?
Gemma-3-4B-IT Abliterated is an experimental uncensored version of Google's Gemma-3-4B model, created using an innovative layerwise abliteration technique. This model represents a significant advancement in AI model modification, specifically designed to reduce content filtering while maintaining core capabilities.
Implementation Details
The model employs a sophisticated layerwise abliteration approach, applying modifications to layers 7 through 29. It uses a unique symmetric pattern of refusal weights ranging from 0.05 to 0.55, resulting in an impressive acceptance rate exceeding 90%. The implementation carefully balances reduced content filtering with output coherence.
- Layerwise abliteration applied across multiple model layers
- Symmetric refusal weight pattern implementation
- Recommended generation parameters: temperature=1.0, top_k=64, top_p=0.95
- Experimental modification technique showing high resilience
Core Capabilities
- High acceptance rate (>90%) for previously filtered content
- Maintained coherent output generation
- Reduced content filtering while preserving model functionality
- Experimental text generation capabilities
Frequently Asked Questions
Q: What makes this model unique?
This model introduces a novel layerwise abliteration technique that differs from traditional approaches by computing refusal directions based on hidden states across multiple layers independently, resulting in more effective content filtering reduction while maintaining output quality.
Q: What are the recommended use cases?
The model is best suited for experimental text generation tasks where reduced content filtering is desired. Users should note the potential for occasional text artifacts (such as minor grammatical inconsistencies) and use the recommended generation parameters for optimal results.
