cua-s1-forms

What it is

A System One model trained for form filling, released under MIT, the most permissive licence in the directory. Where every other entry here chases the four workflows in the community benchmark, this one goes after a task that happens to fit the same shape: a form field is a typed slot, and filling it is a decision from a constrained space, sometimes a closed set, sometimes a bounded number.

Why a narrow model is worth tracking

64 likes on a single-purpose model says the typed-decision interface is being picked up as a general tool rather than as a way to reproduce a benchmark score. That is the more durable outcome for the category, and it is early evidence for it.

It also sets up a test that a general benchmark cannot run. A specialist should beat a generalist on its own task by enough to justify shipping a second model. If it does not, the argument for narrow decision models gets much weaker, and one general model plus a good threshold wins on operational grounds alone.

Status

Not benchmarked, and not comparable on the split we use. Scoring it against LocalLLaMA/typed-decisions would measure it on work it was not built for. A fair test needs a form-filling set with per-field ground truth, which is a different evaluation and one we would build against a real document set rather than a public one.