Phi-3 Safety Post-Training: Aligning Language Models with a "Break-Fix" Cycle

TL;DR

This paper introduces a safety alignment methodology for Phi-3 language models using an iterative 'break-fix' cycle involving dataset curation, safety post-training, benchmarking, and red teaming to enhance responsible AI performance.

Contribution

It presents a novel iterative safety alignment process for small, deployable language models, improving their safety and alignment through multiple rounds of targeted interventions.

Findings

Iterative safety training improved model alignment across benchmarks.

Red teaming identified vulnerabilities and guided safety improvements.

Models demonstrated enhanced safety behavior in multilingual scenarios.

Abstract

Recent innovations in language model training have demonstrated that it is possible to create highly performant models that are small enough to run on a smartphone. As these models are deployed in an increasing number of domains, it is critical to ensure that they are aligned with human preferences and safety considerations. In this report, we present our methodology for safety aligning the Phi-3 series of language models. We utilized a "break-fix" cycle, performing multiple rounds of dataset curation, safety post-training, benchmarking, red teaming, and vulnerability identification to cover a variety of harm areas in both single and multi-turn scenarios. Our results indicate that this approach iteratively improved the performance of the Phi-3 models across a wide range of responsible AI benchmarks. Finally, we include additional red teaming strategies and evaluations that were used to test the safety behavior of Phi-3.5-mini and Phi-3.5-MoE, which were optimized for multilingual capabilities.