kmseong/llama2_7b_chat_only_rsn_tuned_lr5e-5_revised

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:7BQuant:FP8Ctx Length:4kPublished:May 1, 2026License:apache-2.0Architecture:Transformer Open Weights Warm

The kmseong/llama2_7b_chat_only_rsn_tuned_lr5e-5_revised model is a 7 billion parameter Llama-3.2-3B-Instruct variant, fine-tuned by kmseong using the Safety Neuron Tuning (SN-Tune) method. This approach selectively fine-tunes only a small set of 'safety neurons' on the Circuit Breakers dataset, enhancing safety alignment while preserving general capabilities. It is primarily designed for applications requiring improved safety and ethical responses, offering a parameter-efficient method for safety alignment.

Loading preview...

Overview

The kmseong/llama2_7b_chat_only_rsn_tuned_lr5e-5_revised model is a specialized version of the Llama-3.2-3B-Instruct base model, developed by kmseong. It has been fine-tuned using a unique method called Safety Neuron Tuning (SN-Tune), specifically designed to enhance safety alignment without significantly impacting the model's general capabilities. This 7 billion parameter model leverages a context length of 4096 tokens.

Key Capabilities

  • Enhanced Safety Alignment: Utilizes SN-Tune to selectively fine-tune critical 'safety neurons' on the Circuit Breakers dataset.
  • Parameter-Efficient Fine-tuning: Achieves safety improvements by modifying only a small subset of neurons, freezing most other parameters.
  • Preservation of General Capabilities: The SN-Tune method aims to maintain the base model's original performance across various tasks while improving safety.

Good for

  • Applications where improved safety and ethical response generation are critical.
  • Developers looking for a parameter-efficient way to enhance model safety without extensive retraining.
  • Use cases requiring a 7B-class model with a focus on responsible AI outputs.