What it is
jev-curate streams Parquet and JSONL datasets through TypeSafe AI's Jev model to filter synthetic reasoning data, instruction tuning pairs, and web-scraped corpora, emitting clean and rejected records verbatim for dataset engineers.
How it uses Jev
Jev evaluates rows in multi-threaded batches via presets. reasoning-math uses Noul checks (has_circular_logic, is_step_valid) and a Score (reasoning_depth 1-5); anti-sycophancy and code-correctness use Noul checks. Results split rows into clean and rejected outputs.
Primitives:choicescorenoul
Technique worth stealing
Streaming Arrow/Parquet rows through Jev presets with host-side sanity pruning and adaptive token-bucket rate limiting.
Try it
cargo install jev-curate; export TYPESAFE_API_KEY; jev-curate filter train.parquet --preset reasoning-math --out ./output/
Evidence
Each line is one question put to Jev about the README. ≥ 0.60 reads as yes, ≤ 0.40 as no; in between Jev is not making a call.
- Jev-centricyes0.94
- Shows a System One patternyes0.82
- Handles uncertaintyno0.11
- Measuredno0.36
- Runnableyes0.95
- Worth recommendingunclear0.50
- Model replicano0.04
- Problem scopescore on a 0–2 scale1.21
- About Jevyes0.98
Signals by Jev jev-1.13.0, card written by DeepSeek V4.1 Flash from the README on 20 Sept 2026.