Forger: An Environment for Agentic Data Generation

Synthetic data generation has become an essential tool for language model post-training. Synthetic data offers the promise of producing large quantities of high-quality data at less cost than manual data annotation, but naive data generation approaches run into issues with quality, diversity, and generation cost. We describe Forger, a Python library for writing and executing data generation pipelines. Forger casts data generation as programming, focusing on three primitive building blocks for data generation: data acquisition, alteration, and filtering. We show that Forger offers an effective \textit{scaffold} for coding agents, leading to autonomously-generated data that can improve Qwen3-4B on difficult reasoning tasks and outperform an existing data generation library. We show that the agentically-generated data shows excellent diversity, leveraging the provided scaffold.