AutoSynthData: generate validated training tasks for enterprise agents
ServiceNow CoreAI's AutoSynthData converts a target model's failures and a stronger teacher's successes into executable, verified training tasks (system spec, user prompt, verifier) for SFT and iterative curricula.
What changed
ServiceNow CoreAI built AutoSynthData to generate and validate environment-grounded training tasks from model failures using a stronger teacher, target and multiply phases, and sample- and batch-level verification.
Why it matters
Developers and enterprises can get environment-specific, high-quality synthetic tasks to fine-tune agents (SFT) and target persistent capability gaps.
Who should care
DevelopersEnterprisesResearchers
Worth trying
Maybe — If you have an environment, a stronger teacher and the EnterpriseOps Gym released dataset.