kmseong/llama2_7b_chat-WaRP-safeinstr_ratio0.1_lr5e-5

Model Overview

This model, kmseong/llama2_7b_chat-WaRP-safeinstr_ratio0.1_lr5e-5, is a 7 billion parameter Llama 2 based chat model fine-tuned by kmseong. It incorporates a novel Weight space Rotation Process (WaRP) for safety alignment, aiming to create a safer instruction-following model without significantly compromising utility.

Key Capabilities & Features

• Safety-First WaRP Training: Employs a three-phase pipeline (Basis Construction, Importance Scoring, Incremental Learning) to explicitly protect safety mechanisms.
• Gradient Masking: Utilizes gradient masking during fine-tuning to preserve important directions related to safety, ensuring the model maintains refusal capabilities for harmful requests.
• Improved Utility: While prioritizing safety, the model also demonstrates improved utility on reasoning tasks, indicating a balanced approach to safety-utility trade-off.
• Base Model: Built upon the meta-llama/Llama-3.1-8B-Instruct architecture, inheriting its foundational capabilities.
• Dataset Usage: Trained on LibrAI/do-not-answer for safety alignment and openai/gsm8k for utility improvement.

Good For

• Applications requiring enhanced safety alignment in conversational AI.
• Use cases where maintaining refusal capabilities for harmful prompts is critical.
• Scenarios demanding a balance between model utility and safety performance.
• Developers looking for a Llama 2 based model with a specific focus on safety through weight space manipulation.