jev-curate
A Rust command-line tool that filters large synthetic and pretraining text datasets by sending each row through TypeSafe AI's Jev model for calibrated, rubric-based judgments instead of a slower general-purpose language model.
jev-curate is a dataset-curation tool for cleaning large volumes of synthetic reasoning data, instruction-tuning pairs or web-scraped text. It streams Parquet and JSONL data through TypeSafe AI's Jev model, sending every rubric question for a row in one request rather than one round trip per question, and returns typed choice, score and calibrated-probability judgments instead of free-form generated text.
It is distributed as a single Rust binary installed through Cargo, with a command-line interface that filters a dataset file against a named preset, such as one tuned for math reasoning, and writes the kept rows to an output directory. A Python API is also available. The project is an early beta release.