chapter twelve

12 Synthetic data

 

This chapter covers

  • Defining synthetic data
  • Why synthetic data actually works
  • Distilling models from other models
  • Teacher–student knowledge distillation methods
  • Constitutional AI and methods of 
    AI feedback
  • Learning from rubrics and future directions of synthetic data

Reinforcement learning from human feedback is deeply rooted in the idea of maintaining human influence in the models we are building. When the first models were trained successfully with RLHF, human data was the only viable way to improve the models by creating high-quality responses to questions that provided reliable, specific feedback data.

As AI models got better, this assumption rapidly broke down. The possibility of synthetic data, which is far cheaper and easier to iterate on, enabled the proliferation of RLHF by lowering the price of experiments and research. This translated into RLHF being the early center of attention in the broader “post-training” shaping models. This chapter provides a cursory overview of how and why synthetic data is replacing or expanding many pieces of the RLHF pipeline.

12.1 The roles of synthetic data

12.2 Distillation with synthetic data

12.3 The path to on-policy, teacher–student distillation

12.3.1 Adapting knowledge-distillation for LMs

12.3.2 From offline to on-policy distillation

12.3.3 Modern OPD variants

12.4 AI feedback

12.4.1 Balancing AI and human feedback data

12.4.2 Building specific LLMs for judgment

12.5 Constitutional AI

12.5.1 Further reading on CAI

12.6 Rubrics: Prompt-specific AI feedback for training

Summary