kmseong/llama3_2_3b_instruct_MATH_lr5e-5
The kmseong/llama3_2_3b_instruct_MATH_lr5e-5 is a 3.2 billion parameter Llama 3.2 Instruct model, developed by kmseong, that has been Safety Neuron-Tuned (SN-Tune). This fine-tuning method enhances safety alignment by selectively training critical safety neurons on the Circuit Breakers dataset, while preserving general capabilities. It is optimized for improved safety performance in conversational AI applications, offering a more aligned alternative to its base model.
Loading preview...
Overview
The kmseong/llama3_2_3b_instruct_MATH_lr5e-5 is a 3.2 billion parameter instruction-tuned model based on meta-llama/Llama-3.2-3B-Instruct. Its primary differentiator is the application of Safety Neuron Tuning (SN-Tune), a specialized fine-tuning method developed by kmseong. This technique focuses on enhancing the model's safety alignment without significantly impacting its general performance.
Key Capabilities
- Enhanced Safety Alignment: The model has undergone SN-Tune using the Circuit Breakers dataset, specifically targeting and fine-tuning a small set of 'safety neurons'.
- Parameter-Efficient Fine-tuning: SN-Tune freezes most parameters and only fine-tunes the identified safety neurons, making the process efficient.
- Preservation of General Capabilities: This selective tuning approach aims to improve safety while minimizing any degradation of the base model's original instruction-following abilities.
- Llama 3.2 Architecture: Benefits from the robust architecture of the Llama 3.2 Instruct series.
Good For
- Applications requiring a safety-aligned conversational AI model.
- Use cases where mitigating harmful outputs is a priority.
- Developers looking for a Llama 3.2 variant with improved safety characteristics compared to the base model, while maintaining a 3.2B parameter count and a 32768 token context length.