CUA-S1-FORMS
706K-parameter Jev-style option scorer that rates fill, check, click or skip for each form element in one parallel pass.
Overview
cua-s1-forms is a tiny model in the Cua-S1 research family for GUI form filling. Given a UI element and a list of typed options (one per document entity, plus check, click and skip) it returns a probability per option in a single forward pass, without generating text. The README notes this is the same input and output contract as Jev. Every actionable element is scored independently in one batch, and downstream code decides execution order: fills, then checkboxes, then the single submit click.
The architecture is a byte-level embedding with a two-layer Transformer encoder and an option-attention head. Only the safetensors pair is supported; an older pickle checkpoint was removed for security reasons. The model card reports 99.95% on a synthetic test and 100% on a small real demo eval, and the author reported 99.7% against 83.6% for hosted Jev on the same task, noting it is a specialist on its own training domain.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | For this form element, which option applies: a document entity value, check, click or skip? | <entity values> · check · click · skip |
How it works
- Snapshot the form's actionable elements
- Score every element against the options in one batch
- Code orders fills, then checks, then the submit click
- Execute through Cua Driver
Strengths
- Very small (about 2.8 MB) and fast
- Scores all elements in parallel
- Signed safetensors checkpoint validated on load
Caveats
- Specialist model; comparison with Jev is on its own domain
- Research release
- Requires the cua-s1 package from the Cua repository
Getting started
hf download cua-ai/cua-s1-forms --exclude "*.pt" --local-dir ./cua-s1-forms, then load with cua_s1.model.load_checkpoint.At a glance
- Type
- Model
- Status
- research
- Language
- Python
- Runtime
- PyTorch model
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Specialist decision model: a 706,048-parameter, 2.8 MB jev-like option scorer that rates FILL / CHECK / CLICK / SKIP for each form field in one parallel pass, reporting 99.7% on its own form-filling eval against Jev's 83.6% - a specialist on home turf rather than a general win.
yibie/awesome-jev
Frequently asked questions
How does it relate to Jev?
It uses the same contract of options in, one probability per option out, but is a tiny specialist model.
Is it better than Jev?
The author reported 99.7% versus 83.6% for hosted Jev on this form task, which the specialist was trained on.
Why only safetensors?
A pickle checkpoint was removed because unpickling can run arbitrary code.
Related projects
Computer-use platform (Cua Spaces, Driver, Lume) with CUA-S1 small decision models and a preview jev-use example for bounded actions.
ProjectbetaPythonFrom-scratch research reconstruction of a small System-One decision model, used to route and gate commands in a Windows palette.
ProjectresearchPythonOneJev: open multimodal System One model (0.8B-27B) answering typed questions about screenshots, photos, video and text in one pass.
ModelbetaPythonOpenJev 1.7B-8B: open decision models distilled for general decisions and browser-agent steps, served behind /v1/systemone.
ModelbetaPythonOpenJev on Hugging Face: open-weights decision model for text, web pages and screenshots with up to 52 options per forward pass.
ServicebetaPython- CLM★ 3,014
CLM-8B, a contrastively trained System One model that embeds states and actions separately, served behind a TypeSafe-compatible API.
ServiceresearchPython