DataMaster: Autonomous Agent for End-to-End Data Engineering in ML Pipelines

arXiv·medium signal

DataMaster studies task-conditioned autonomous data engineering where an agent searches for external datasets, adapts them to existing pipelines, validates candidates through downstream training, and carries forward lessons from prior attempts. The system automates the increasingly bottleneck phase of ML — data curation — as model architectures and training recipes become standardized. Demonstrates that autonomous data engineering can match human data scientists on standard ML benchmarks.