kmseong/llama2_7b-chat-Safety-FT-lr5e-5

Model Overview

The kmseong/llama2_7b-chat-Safety-FT-lr5e-5 is a 7 billion parameter model built upon the Llama 2 architecture, specifically fine-tuned for safety alignment in chat-based applications. This model differentiates itself through its training methodology, which involves applying MLP to the attention mechanism's q, k, and v components, alongside per-layer adjustments. The training process then proceeds with a non-freeze approach, allowing for comprehensive adaptation.

Key Capabilities

• Safety Alignment: Primary focus on enhancing safety in conversational AI.
• Llama 2 Base: Leverages the robust Llama 2 7B architecture.
• Custom Training: Utilizes specific modifications to attention and MLP layers during fine-tuning.
• Context Length: Supports a context window of 4096 tokens.

Good For

• Developing safer conversational agents.
• Research into safety alignment techniques for large language models.
• Applications requiring a Llama 2-based model with enhanced safety characteristics.

Citation

This model is associated with the following citation:

@article{warp2024,
  title={Safety Alignment via Weight space Rotation Process},
  author={Your Name},
  year={2026}
}