Balancing Enhancement, Harmlessness, and General Capabilities: Enhancing Conversational LLMs with Direct RLHF

TL;DR

This paper introduces a novel approach to improve conversational LLMs by directly applying Harmless RLHF, bypassing traditional supervised fine-tuning, which preserves capabilities and reduces toxicity, demonstrated on the Mistral model.

Contribution

The paper proposes a direct RLHF method that enhances conversational abilities and safety of LLMs without the drawbacks of supervised fine-tuning, applied to the Mistral model.

Findings

Mistral-Plus outperforms similar-sized open-source models.

Significant improvement in conversational abilities.

Reduced toxic output generation.

Abstract

In recent advancements in Conversational Large Language Models (LLMs), a concerning trend has emerged, showing that many new base LLMs experience a knowledge reduction in their foundational capabilities following Supervised Fine-Tuning (SFT). This process often leads to issues such as forgetting or a decrease in the base model's abilities. Moreover, fine-tuned models struggle to align with user preferences, inadvertently increasing the generation of toxic outputs when specifically prompted. To overcome these challenges, we adopted an innovative approach by completely bypassing SFT and directly implementing Harmless Reinforcement Learning from Human Feedback (RLHF). Our method not only preserves the base model's general capabilities but also significantly enhances its conversational abilities, while notably reducing the generation of toxic outputs. Our approach holds significant implications for fields that demand a nuanced understanding and generation of responses, such as customer service. We applied this methodology to Mistral, the most popular base model, thereby creating Mistral-Plus. Our validation across 11 general tasks demonstrates that Mistral-Plus outperforms similarly sized open-source base models and their corresponding instruct versions. Importantly, the conversational abilities of Mistral-Plus were significantly improved, indicating a substantial advancement over traditional SFT models in both safety and user preference alignment.