kmseong/llama2_7b-chat-Safety-FT-lr3e-5

Overview

The kmseong/llama2_7b-chat-Safety-FT-lr3e-5 is a 7 billion parameter language model built upon the Llama 2 architecture. This model has undergone a specific fine-tuning process focused on safety alignment, distinguishing it from general-purpose Llama 2 chat models. The training methodology involves applying attention mechanisms (q, k, v) and MLP layers (up, down), along with a 'perlayer' application, all within a non-freeze training paradigm.

Key Capabilities

• Safety Alignment: Specifically fine-tuned to improve safety characteristics in conversational AI.
• Llama 2 Base: Leverages the robust architecture of the Llama 2 7B model.
• Non-Freeze Training: Employs a training approach where model layers are not frozen, allowing for comprehensive adaptation during fine-tuning.
• Targeted Layer Application: Incorporates specific modifications to attention and MLP layers during the fine-tuning process.

Good For

• Developing chat applications where safety and alignment are primary concerns.
• Researching and implementing safety alignment techniques in large language models.
• Building conversational agents that require a specialized safety-focused fine-tuning over a Llama 2 base.

Citation

This model is associated with research on "Safety Alignment via Weight space Rotation Process". For more details, refer to the citation:

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