AutoSynthData is a framework designed to automate the generation of synthetic training data for agentic models by focusing on environment-specific capability gaps. By leveraging a stronger teacher model to generate and validate tasks based on target model failures, the system creates a curriculum that iteratively improves agent performance within constrained enterprise environments.
Key points
Effective agent training requires tasks that are feasible within the environment, representative of real requests, and programmatically verifiable.
Targeted synthetic data generation can be automated by using a stronger teacher model to synthesize training examples based on a smaller model's failure points.
The training curriculum should dynamically shift to focus on remaining capability gaps as the target model improves.
Validated synthetic data can significantly improve Pass@1 metrics in complex, multi-domain environments like ITSM and Hybrid Ops.
