kmseong/qwen2_5_32b_instruct_gsm8k_full_finetune_after_ssft_lr5e-5

Model Overview

This model, kmseong/qwen2_5_32b_instruct_gsm8k_full_finetune_after_ssft_lr5e-5, is a specialized version of the meta-llama/Llama-3.2-3B-Instruct base model. It features 32.8 billion parameters and a context length of 32768 tokens.

Key Capabilities & Differentiators

• Safety Neuron Tuning (SN-Tune): This model utilizes a unique fine-tuning approach where only a small set of "safety neurons" are identified and fine-tuned on safety alignment data (Circuit Breakers dataset). All other parameters remain frozen.
• Enhanced Safety Alignment: The primary goal of SN-Tune is to significantly improve the model's safety characteristics compared to its base model.
• Parameter-Efficient Fine-tuning: By selectively fine-tuning only critical neurons, the SN-Tune method aims to achieve safety improvements with minimal computational overhead and without compromising the model's general capabilities.

Ideal Use Cases

This model is particularly well-suited for applications where:

• Robust safety alignment is paramount: Users require a model with improved resistance to generating unsafe or undesirable content.
• Maintaining general instruction-following abilities is important: The SN-Tune method is designed to enhance safety without degrading the model's broader performance.
• Efficient fine-tuning is a consideration: The selective tuning approach offers a resource-effective way to achieve safety improvements.