CUA-Lite solves the fragmentation that blocks teams from building and testing computer-use agents. Today agents, environments, trace data and training frameworks live in separate repositories with incompatible interfaces, forcing engineers to write glue code for every combination. CUA-Lite puts all four pieces behind a single action space, a shared LiteSample data schema and one command line interface, so the same code works on desktop, browser and mobile sandboxes.
The biggest practical barrier is the virtual‑machine tax. Traditional OSWorld benchmarks require a full QEMU/KVM VM per task, which needs nested virtualization, consumes over 4 GB of RAM and starts slowly. Lite.OSWorld reproduces the exact same task suite inside a plain Docker container, using only 0.9 GB per instance and launching in under 24 seconds. On a typical host this yields roughly 4.6 times more parallel desktops without sacrificing fidelity—scores across 13 models match the VM baseline, meaning any training signal earned in the container transfers directly to the real benchmark.
Data handling is unified through LiteSample, a supervised‑learning schema stored as Parquet plus images. More than ten existing CUA datasets have been pre‑processed into this format and are freely available, together with rollout data generated by frontier teacher models for distillation. Per‑model adapters convert LiteSample into each framework’s expected format, eliminating custom preprocessing pipelines.
Training and evaluation share the same lite.gym loop: screenshots go in, actions come out, with one action space per platform. Swapping a model ID or environment ID in a single script runs rollouts, supervised fine‑tuning or reinforcement learning. A reported fine‑tune of Qwen3‑VL‑2B‑Instruct on Lite.ScaleCUA lifted episode return from 0.138 to 0.237 on the Lite.OSWorld split using two GPUs, showing measurable gains with minimal setup.
The platform is deployable on any Docker host, requires no special kernel access, and supports desktop, browser and mobile environments in a unified workflow.
#AI #MachineLearning #OpenSource #Automation #DevOps #ComputerVision