Why CUA-Lite — and a call for contributors.
Why CUA-Lite
Computer-use agent resources are fragmented. Sandboxes are heavy and unstandardized: a virtual machine per task, each with its own interface and action space. Every dataset picks its own schema, so data collected for one agent can't train another. And no framework standardizes how computer-use agents are evaluated and trained, so every project rebuilds the same plumbing — agent loops, VM setup, eval scripts — before a single rollout runs.
CUA-Lite puts all of it in one place — all a computer-use agent needs, open to share, simple to run: (1) sandboxes with verifiable tasks for benchmarking and training at scale; (2) unified SFT data, 10+ datasets and fresh rollouts, free on Hugging Face — all through one framework to eval, SFT, and RL any agent. Here's how each compounds:
Sandboxes & verifiable tasks
Lightweight, KVM-free sandboxes pack many to a machine, each with verifiable tasks. They run in parallel — and each task serves both training and benchmarking.
Call for sandbox contributors. A sandbox only matters while people run it. Add yours, and every agent trains and benchmarks on it — now and later. One integration, and the whole field builds on it.
Datasets
Convert a dataset once, and every agent can train on it.
LiteSample is the one schema, shared
across every env, agent, and task type, free on Hugging Face.
LiteSample{ }CUA-Lite ships an adapter per model,
packing a unified LiteSample into the exact
training format each one needs. The figure above shows one, with the building
blocks to add your own. The next figure plays out Qwen3.5's.
Call for data contributors. Data only matters while models can train on it. Share yours, and every agent trains on it — even models that don't exist yet. One conversion, and the whole community trains on it.
Preprocessing guide ↗ · Agent harnesses ↗ · SFT guide ↗
One framework: eval & RL
Any agent plugs into any environment through one interface.
They meet in lite.gym — screenshots up,
actions down, in one action space plus each env's extra tools. The same loop
serves both eval and RL: a rollout is scored to rank an agent, or trained on to
improve it.
lite.gymBring a dataset, an env, or an agent — each one compounds. Or just tell us what's missing — GitHub · Hugging Face · Email.