Data-driven control of the radio access network through O-RAN RAN Intelligent Controllers depends on labelled, stack-wide measurement data, which is scarce and costly to obtain on commercial equipment. This paper describes a fully software-defined 5G Standalone (SA) testbed and an automated pipeline that turns it into a generator of labelled machine-learning datasets. The testbed integrates the srsRAN gNodeB, the Open5GS core, a GNU Radio plus ZeroMQ radio emulator with runtime-controllable channel impairment, and a FlexRIC Near-RT RIC, all on a single commodity server with CPU isolation and per-user network namespaces for repeatability. A Monte-Carlo orchestrator runs up to one thousand 20-second iterations, randomizing user count, per-user path loss, traffic type and rate, and parallelism, and a synchronized parser assembles one-second snapshots into a 37-feature table spanning the PHY, MAC, RLC/PDCP, and operating-system layers, augmented with five-second rolling statistics, an energy-efficiency index, and forward-looking regression and classification targets at a five-second horizon. We describe the architecture, the feature engineering, the labelling methodology, and the resulting dataset structure, and we discuss how the dataset supports predictive xApps for proactive congestion and degradation control. The platform and pipeline are designed for reproducibility and are intended as a reusable resource for the O-RAN machine-learning community.