kmseong/llama31_8b_instruct_math_ft_freeze_sn_lr1e-5

TEXT GENERATIONConcurrency Cost:1Model Size:8BQuant:FP8Ctx Length:32kPublished:Apr 18, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

The kmseong/llama31_8b_instruct_math_ft_freeze_sn_lr1e-5 is an 8 billion parameter Llama-3.2-3B-Instruct model fine-tuned by kmseong with a 32768 token context length. It utilizes a Safety Neuron Tuning (SN-Tune) method, which selectively fine-tunes only safety-critical neurons on safety alignment data. This approach enhances safety while preserving general capabilities, making it suitable for applications requiring improved safety alignment.

Loading preview...

Overview

This model, kmseong/llama31_8b_instruct_math_ft_freeze_sn_lr1e-5, is an 8 billion parameter instruction-tuned variant of the meta-llama/Llama-3.2-3B-Instruct base model. It has been specifically fine-tuned using a novel method called SN-Tune (Safety Neuron Tuning) to enhance its safety alignment.

Key Capabilities & Features

  • Safety Neuron Tuning (SN-Tune): A selective fine-tuning approach that identifies and targets only a small set of "safety neurons" critical for alignment.
  • Parameter-Efficient Fine-tuning: By freezing all non-safety parameters and only fine-tuning safety neurons on the Circuit Breakers dataset, the model achieves enhanced safety with minimal impact on its general capabilities.
  • Improved Safety Alignment: Designed to offer better safety performance compared to its base model, making it suitable for sensitive applications.
  • Llama-3.2-3B-Instruct Base: Benefits from the strong foundational capabilities of the Llama-3.2-3B-Instruct architecture.

When to Use This Model

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

  • Enhanced safety alignment is a primary concern.
  • Maintaining general model capabilities while improving safety is crucial.
  • Applications require a parameter-efficient approach to safety fine-tuning.

It is licensed under the Apache 2.0 License, inheriting terms from its base model.