kmseong / llama3_2_3b-instruct-SSFT-lr5e-5

WaRP-Safety-Llama3_8B_Instruct

Fine-tuned Llama 3.1 8B Instruct model for safety alignment using Weight space Rotation Process (WaRP).

Model Details

• Base Model : meta-llama/Llama-3.1-8B-Instruct
• Training Method : Safety-First WaRP (3-Phase pipeline)
• Training Date : 2026-04-28

Training Procedure

Phase 1: Basis Construction

• Collected activations from FFN layers using safety data
• Computed SVD to obtain orthonormal basis vectors
• Identified 419 important neurons in layer 31

Phase 2: Importance Scoring

• Calculated importance scores using gradient-based methods
• Generated masks for important directions
• Used teacher forcing on safety responses

Phase 3: Incremental Learning

• Fine-tuned on utility task (GSM8K) with gradient masking
• Protected important directions to maintain safety
• Improved utility while preserving safety mechanisms

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "kmseong/WaRP-Safety-Llama3_8B_Instruct"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto")

# Generate text
inputs = tokenizer("What is machine learning?", return_tensors="pt")
outputs = model.generate(**inputs, max_length=100)
print(tokenizer.decode(outputs[0]))

Safety Features

• ✅ Protected safety mechanisms through gradient masking
• ✅ Maintained refusal capability for harmful requests
• ✅ Improved utility on reasoning tasks
• ✅ Balanced safety-utility tradeoff

Datasets

• Safety Data : LibrAI/do-not-answer
• Utility Data : openai/gsm8k

Citation

@article{warp-safety,
  title={Safety-First WaRP: Weight space Rotation Process for LLM Safety Alignment},
  author={Min-Seong Kim},
  year={2026}
}

License

This model is built on Llama 3.1 8B Instruct and follows the same license.

Disclaimer

This model is fine-tuned for improved safety. Users should evaluate model outputs for their specific use cases and apply additional safety measures as needed.