jev-curate

What it is

High-Throughput Synthetic & Pretraining Dataset Sifter Powered by TypeSafe AI (Jev)

The homepage listed on the repository, jev-curate.vercel.app, puts it like this: Curate synthetic and pretraining datasets in Rust with TypeSafe Jev judgment. Parquet and JSONL I/O, speculative fan-out, unmetered output, real benchmarks.

What the README says it does

• Generative LLMs are too slow and expensive: Running Claude 3.5 Sonnet or GPT-4o to judge synthetic rows costs $15,000–$50,000 per billion tokens and crawls at a painful 30–50 rows/sec.
• Regex heuristics are blind to reasoning flaws: Keyword and regex filters can check syntax, but fail to detect circular reasoning, hallucinated derivation steps, or robotic sycophancy.
• Context rot from uncompressed inputs: Naively feeding raw data into LLMs causes decision accuracy to crater while burning money on boilerplate text.
• $4.20 per 100M tokens: Jev bills $0.042/Mtok for input, zero for output. TypeSafe benchmarks System One workflows 444.6x cheaper and 193.6x faster than generative LLMs (source).
• Mathematical calibration: Receives calibrated probabilities (Noul), ordinal rubrics (Score on a 0 to 4 scale, sent as an ordered list), and categorical choices (Choice), eliminating generative text slop.

From the project’s README.

Where the decision sits

jev-curate calls the hosted decision API rather than running its own weights. That keeps the project small and means its behaviour moves whenever the vendor ships a new model version. It also means an API key and a network round trip sit in the path of every decision.

Models it references: Jev.

How active it is

Created 2026-09-18, last pushed 2026-10-02, 101 stars and 16 forks when we read it on 2026-11-02. Everything in this category is weeks old at most, so a stale push date means something different here than it would on a mature project. Check the repository before you depend on it.