kmseong/llama2_7b-SSFT-WaRP_agnews_FT_lr3e-5

Model Overview

This model, kmseong/WaRP-Safety-Llama3_8B_Instruct, is a fine-tuned version of the meta-llama/Llama-3.1-8B-Instruct base model, developed by Min-Seong Kim. It leverages a novel Safety-First Weight space Rotation Process (WaRP), a 3-phase pipeline designed to enhance safety alignment in large language models.

Key Capabilities & Features

• Enhanced Safety Alignment: Utilizes a sophisticated WaRP method to protect safety mechanisms through gradient masking, ensuring robust refusal capabilities for harmful requests.
• Balanced Safety-Utility Tradeoff: Achieves an improved balance between safety and utility, demonstrating enhanced performance on reasoning tasks (e.g., GSM8K) while preserving safety.
• Targeted Fine-tuning: The training procedure involved identifying and protecting important neuronal directions, allowing for incremental learning on utility tasks without compromising safety.

Training Details

The WaRP training involved three phases:

1. Basis Construction: Collected FFN layer activations from safety data (LibrAI/do-not-answer) to compute orthonormal basis vectors.
2. Importance Scoring: Used gradient-based methods to score and mask important directions, guided by teacher forcing on safety responses.
3. Incremental Learning: Fine-tuned on utility data (openai/gsm8k) with gradient masking to improve utility while safeguarding critical safety directions.

Ideal Use Cases

This model is particularly well-suited for applications where:

• Safety is paramount: Requiring a model that reliably refuses harmful or inappropriate queries.
• Reasoning tasks are important: Benefiting from improved utility on tasks like mathematical problem-solving.
• Balancing safety and performance is a key requirement, offering a robust solution for general-purpose conversational AI with strong safety guardrails.