llama3_2_3b_instruct_MATH_lr5e-5

llama3_2_3b_instruct_MATH_lr5e-5

This is a Safety Neuron-Tuned (SN-Tune) version of Llama-3.2-3B-Instruct.

Model Description

• Base Model: meta-llama/Llama-3.2-3B-Instruct
• Fine-tuning Method: SN-Tune (Safety Neuron Tuning)
• Training Data: Circuit Breakers dataset (safety alignment data)
• Upload Date: 2026-05-02 19:28:53

What is SN-Tune?

SN-Tune is a selective fine-tuning approach that:

1. Detects safety neurons - a small set of neurons critical for safety
2. Freezes all non-safety parameters
3. Fine-tunes only safety neurons on safety data

This approach allows for:

• Enhanced safety alignment
• Minimal impact on general capabilities
• Parameter-efficient fine-tuning

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "kmseong/llama3_2_3b_instruct_MATH_lr5e-5"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

# Generate text
prompt = "How can I help you today?"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_length=100)
print(tokenizer.decode(outputs[0]))

Safety Note

This model has been fine-tuned specifically for safety using the SN-Tune method. It should provide improved safety alignment compared to the base model.

License

This model is licensed under the Apache 2.0 License. See the base model (meta-llama/Llama-3.2-3B-Instruct) for more details.

References

• Base model: meta-llama/Llama-3.2-3B-Instruct
• Safety neurons detection methodology