ZySec-AI/gemma-3-27b-tools: Abliterated Gemma 3 27B with Tool Calling
This model, developed by ZySec-AI, is an experimental, uncensored variant of Google's Gemma 3 27B IT. It introduces a novel "abliteration" technique designed to remove refusal behaviors while maintaining the model's inherent capabilities. The primary focus of its development has been to improve output coherence and effectively handle diverse requests through minimal fine-tuning.
Key Capabilities
- Tool Calling Enabled: The model is equipped with tool support, significantly enhancing its flexibility and applicability across a wider range of tasks requiring external function calls.
- Refusal Removal: Utilizes an "abliteration" technique that computes refusal directions based on independent hidden states for each layer, effectively mitigating unwanted refusals.
- Enhanced Output Coherence: Minimal fine-tuning has been applied to specifically improve the consistency and quality of the model's generated outputs.
Recommended Usage
For optimal performance, the following generation parameters are recommended:
temperature=1.0top_k=64top_p=0.95
What Makes This Model Different?
Unlike standard instruction-tuned models, ZySec-AI/gemma-3-27b-tools stands out due to its "abliteration" process, which specifically targets and removes refusal mechanisms. This makes it a unique option for use cases where unfiltered responses and direct task execution via tool calling are prioritized, without extensive re-training.