Synthetic Data for any Differentiable Target

TL;DR

This paper introduces Dataset Policy Gradient (DPG), a reinforcement learning method that optimizes synthetic data generators to produce targeted examples, effectively controlling language model behaviors through synthetic training data.

Contribution

The paper presents DPG, a novel RL primitive that precisely optimizes synthetic data for targeted model behavior, enabling complex control without explicit prompts.

Findings

Targeted embedding of QR codes, patterns, and UUIDs in language models.

Ability to rephrase inputs in new languages using synthetic data.

Reduction of model's $ ext{l}^2$ norm through synthetic fine-tuning.

Abstract

What are the limits of controlling language models via synthetic training data? We develop a reinforcement learning (RL) primitive, the Dataset Policy Gradient (DPG), which can precisely optimize synthetic data generators to produce a dataset of targeted examples. When used for supervised fine-tuning (SFT) of a target model, these examples cause the target model to do well on a differentiable metric of our choice. Our approach achieves this by taking exact data attribution via higher-order gradients and using those scores as policy gradient rewards. We prove that this procedure closely approximates the true, intractable gradient for the synthetic data generator. To illustrate the potential of DPG, we show that, using only SFT on generated examples, we can cause the target model's LM head weights to (1) embed a QR code, (2) embed the pattern $\texttt{67}$, and (3) have lower $\ell^2$ norm. We additionally show that we can cause the generator to (4) rephrase inputs in a new language and (5) produce a specific UUID, even though neither of these objectives is conveyed in the generator's input prompts. These findings suggest that DPG is a powerful and flexible technique for shaping model properties using only synthetic training examples.