wvnvwn/gemma-2-9b-it-gsm8k-sn-tuned-lr3e-5

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:9BQuant:FP8Ctx Length:16kPublished:May 3, 2026License:apache-2.0Architecture:Transformer Open Weights Warm

The wvnvwn/gemma-2-9b-it-gsm8k-sn-tuned-lr3e-5 model is a 9 billion parameter instruction-tuned language model, based on meta-llama/Llama-3.2-3B-Instruct, with a 16384 token context length. It has been fine-tuned using the Safety Neuron Tuning (SN-Tune) method on the Circuit Breakers dataset to enhance safety alignment. This model is specifically optimized for improved safety performance while preserving general capabilities through parameter-efficient fine-tuning.

Loading preview...

Overview

This model, wvnvwn/gemma-2-9b-it-gsm8k-sn-tuned-lr3e-5, is a 9 billion parameter instruction-tuned variant of the meta-llama/Llama-3.2-3B-Instruct base model. Its primary distinguishing feature is the application of Safety Neuron Tuning (SN-Tune), a selective fine-tuning approach designed to enhance safety alignment.

Key Capabilities & Features

  • Enhanced Safety Alignment: Fine-tuned specifically to improve safety performance.
  • Parameter-Efficient Fine-tuning: Utilizes SN-Tune, which identifies and fine-tunes only a small subset of "safety neurons" while freezing other parameters.
  • Preservation of General Capabilities: The SN-Tune method aims to minimize impact on the model's broader abilities.
  • Base Model: Built upon the robust Llama-3.2-3B-Instruct architecture.

What is SN-Tune?

SN-Tune is a methodology that involves:

  1. Detecting specific "safety neurons" within the model.
  2. Freezing all non-safety-critical parameters.
  3. Fine-tuning only these identified safety neurons using safety-specific datasets, such as the Circuit Breakers dataset.

Use Cases

This model is particularly suitable for applications where safety and responsible AI deployment are paramount. Developers looking for a Llama-3.2-3B-Instruct-based model with improved safety characteristics, without significantly compromising general performance, should consider this version.