kmseong/llama2_7b_base_resta_lr3e-5_y0.3

TEXT GENERATIONConcurrency Cost:1Model Size:7BQuant:FP8Ctx Length:4kPublished:Apr 23, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

The kmseong/llama2_7b_base_resta_lr3e-5_y0.3 is a 7 billion parameter Llama 2 base model, specifically a Safety Neuron-Tuned (SN-Tune) version of Llama-3.2-3B-Instruct. This model utilizes a selective fine-tuning approach to enhance safety alignment by only fine-tuning critical safety neurons on the Circuit Breakers dataset, while minimizing impact on general capabilities. It is designed for applications requiring improved safety alignment with parameter-efficient fine-tuning.

Loading preview...

Overview

This model, kmseong/llama2_7b_base_resta_lr3e-5_y0.3, is a 7 billion parameter variant of the Llama 2 architecture, specifically fine-tuned from meta-llama/Llama-3.2-3B-Instruct. Its core innovation lies in the application of SN-Tune (Safety Neuron Tuning), a method developed to enhance model safety without compromising general performance.

Key Capabilities

  • Enhanced Safety Alignment: Achieved through selective fine-tuning of 'safety neurons' on the Circuit Breakers dataset.
  • Parameter-Efficient Fine-tuning: Only a small subset of critical neurons are fine-tuned, leaving most parameters frozen.
  • Minimal Impact on General Capabilities: The SN-Tune method aims to preserve the base model's original performance while improving safety.

What is SN-Tune?

SN-Tune is a targeted fine-tuning approach that identifies and isolates a small set of neurons crucial for safety. It then freezes all other parameters and exclusively fine-tunes these safety neurons using dedicated safety data. This allows for efficient safety improvements.

Good For

  • Applications where safety and responsible AI behavior are paramount.
  • Scenarios requiring a Llama 2-based model with improved safety alignment over its base counterpart.
  • Use cases benefiting from parameter-efficient fine-tuning for safety.

License

This model operates under the Apache 2.0 License, inheriting terms from its base model, meta-llama/Llama-3.2-3B-Instruct.