wvnvwn/qwen2.5-7b-instruct-gsm8k-sn-tuned-lr5e-5

TEXT GENERATIONConcurrency Cost:1Model Size:7.6BQuant:FP8Ctx Length:32kPublished:May 2, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

The wvnvwn/qwen2.5-7b-instruct-gsm8k-sn-tuned-lr5e-5 is a 7.6 billion parameter instruction-tuned causal language model, based on the Llama-3.2-3B-Instruct architecture, with a 32K context length. It has been fine-tuned using the Safety Neuron Tuning (SN-Tune) method on safety alignment data. This model is specifically designed to enhance safety alignment while preserving general capabilities through parameter-efficient fine-tuning.

Loading preview...

Overview

This model, wvnvwn/qwen2.5-7b-instruct-gsm8k-sn-tuned-lr5e-5, is a 7.6 billion parameter instruction-tuned variant of the Llama-3.2-3B-Instruct base model. Its primary distinction lies in its fine-tuning approach: Safety Neuron Tuning (SN-Tune). This method selectively fine-tunes only a small set of "safety neurons" on dedicated safety alignment data, specifically the Circuit Breakers dataset, while keeping all other parameters frozen.

Key Capabilities & Features

  • Enhanced Safety Alignment: Significantly improves the model's safety profile compared to its base model.
  • Parameter-Efficient Fine-tuning: Achieves safety improvements with minimal impact on the model's general capabilities, as only a fraction of parameters are adjusted.
  • Llama-3.2-3B-Instruct Base: Inherits the foundational capabilities of the Llama-3.2-3B-Instruct architecture.

When to Use This Model

This model is particularly well-suited for applications where:

  • Safety and ethical considerations are paramount: Its SN-Tune fine-tuning makes it a strong candidate for sensitive use cases.
  • Maintaining general performance is crucial: The SN-Tune method aims to enhance safety without degrading broader language understanding and generation abilities.
  • Resource efficiency is a concern: The parameter-efficient fine-tuning approach is beneficial for deployment and further customization.