ChatGLM-RLHF: Practices of Aligning Large Language Models with Human Feedback
This work addresses alignment challenges for users of ChatGLM, but it is incremental as it builds on existing RLHF methods with specific optimizations.
The authors tackled the problem of aligning large language models with human preferences by developing the ChatGLM-RLHF pipeline, which improved alignment by achieving an average 15% more wins against the supervised fine-tuned version in Chinese tasks.
ChatGLM is a free-to-use AI service powered by the ChatGLM family of large language models (LLMs). In this paper, we present the ChatGLM-RLHF pipeline -- a reinforcement learning from human feedback (RLHF) system -- designed to enhance ChatGLM's alignment with human preferences. ChatGLM-RLHF encompasses three major components: the collection of human preference data, the training of the reward model, and the optimization of policies. Throughout the process of integrating ChatGLM-RLHF into production, we encountered and addressed several unprecedented challenges. We introduce the strategies to mitigate reward variance for stabilized large-scale training, implement model parallelism with fused gradient-descent, and design regularization constraints to avoid catastrophic forgetting in LLMs. Experiments show that ChatGLM-RLHF brings significant improvements in alignment tasks compared to the supervised fine-tuned (SFT) version of ChatGLM. For instance, it achieves on average 15\% more wins against ChatGLM-SFT in Chinese alignment tasks. The work presents our practices of aligning LLMs with human preferences, offering insights into the challenges and solutions in RLHF implementations.