Our Work

Benchmarks, Evals,
and Datasets

Research on AI capabilities in real-world workflows. We build rigorous evaluations and training environments for frontier models.

Datasets

Jan 7, 2026

Gauntlet

An SFT dataset of real developer workflows for training production-grade code models.

Benchmarks

Jan 6, 2026

PDF Bench

We tested three frontier AI models on 70 real-world PDF form-filling tasks. The results were more surprising, and more sobering, than we expected.

We work with frontier labs to provide evals and environments that are economically relevant and non reward hackable. Email us at founders@refresh.dev to get in touch.

Benchmarks, Evals,Benchmarks, Evals, andand DatasetsDatasets

Gauntlet

PDF Bench

Benchmarks, Evals,
and Datasets