kmseong/llama-3.2-3b-instruct-only-rsn-tuned-lr5e-5

Model Overview

This model, llama-3.2-3b-instruct-only-rsn-tuned-lr5e-5, is a 3.2 billion parameter instruction-tuned variant of the meta-llama/Llama-3.2-3B-Instruct base model. Developed by kmseong, its primary distinction lies in its fine-tuning methodology: Safety Neuron Tuning (SN-Tune).

Key Capabilities & Features

• Enhanced Safety Alignment: The model has undergone SN-Tune, a selective fine-tuning process that identifies and adjusts only a small subset of 'safety neurons' critical for responsible AI behavior.
• Parameter-Efficient Fine-tuning: By freezing most parameters and only fine-tuning safety-critical neurons, this method achieves safety improvements with minimal computational overhead.
• Preservation of General Capabilities: The SN-Tune approach aims to enhance safety without significantly impacting the model's broader instruction-following and general reasoning abilities.
• Specialized Training Data: Fine-tuned on the "Circuit Breakers" dataset, specifically designed for safety alignment.

When to Use This Model

This model is particularly well-suited for use cases where:

• Safety and responsible AI are paramount: It offers improved safeguards against generating harmful or undesirable content compared to its base model.
• Maintaining general instruction-following is important: The SN-Tune method is designed to minimize degradation of non-safety-related performance.
• Resource efficiency is a concern: The parameter-efficient fine-tuning makes it a practical choice for deploying safety-aligned LLMs.