Jev Curate

A target. The 24.0 rows/sec figure comes from a local mock bench (examples/bench_mock.rs) on one node, and the README labels v0.2.0 beta. Real single-node throughput is capped by TypeSafe's 1,200 requests/minute, roughly 20 rows/sec, because each row is one HTTP request.

No. Those come from TypeSafe's own benchmarks with the source linked to typesafe.ai, not from jev-curate. The only measurement the repo makes itself is the mock-bench throughput, and its own tests run against an in-process mock server with zero live credits.

Verbatim, always. Rewriting is the first listed non-goal: it never paraphrases or regenerates text, so formulas and math survive untouched. Output is clean.jsonl, rejected.jsonl, errors.jsonl and a manifest, with Parquet opt-in via --format parquet to keep the schema.

Both have a path. --dry-run does host pre-filtering only and needs no TYPESAFE_API_KEY, and pip install jev-curate works from manylinux, macOS and Windows wheels. The Python API is the exception: PyJevCurator comes from maturin develop on a clone with the python cargo feature enabled.

Install the Rust CLI with `cargo install jev-curate`, export TYPESAFE_API_KEY, then point it at a file: `jev-curate filter train.parquet --preset reasoning-math --out ./output/ --concurrency 32`. Python users can `pip install jev-curate` from the prebuilt manylinux, macOS and Windows wheels instead.

It uses Jev Noul and Score rubrics together: the reasoning-math preset checks circular logic and step validity with yes/no questions and grades reasoning depth on a 1-5 score. Every rubric question for a row is evaluated in one HTTP request via speculative parallel fan-out.

Failed rows are written to errors.jsonl, kept separate from clean.jsonl and rejected.jsonl, so a failed API call does not get silently dropped into either outcome file. A manifest accompanies the three outputs to describe the run.