jev-curate

Support: fuel the next build -

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

Beta: v0.2.0 is experimental. Expect rough edges. Please contribute by opening an issue or PR.

Live: jev-curate.vercel.app

Measured: 23.9 rows/sec single-node local mock (bash bench/run_mock.sh, results in bench/results/mock.json). Not measured: 1,500+ rows/sec is a horizontal-scaling target, not a benchmark result.

Product Hunt: jev-curate Upcoming (launches Oct 13, 12:01am PT)

By Akash Priyadarshi

Why jev-curate • Quickstart • CLI Reference • Python API • Architecture • Non-Goals • Ecosystem

---

Why jev-curate?

Cleaning 10M to 1B rows of synthetic reasoning data, instruction tuning pairs, or web-scraped corpora is an economic and technical nightmare:

• 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.

jev-curate solves this by piping Parquet/JSONL through TypeSafe AI's Jev model (jev-latest, pin with JEV_MODEL=jev-1.13.0):

• Single-round-trip rubrics per row: Evaluates all rubric questions concurrently in one HTTP request per row via Jev's speculative parallel fan-out with zero per-question round-trips. Single-node throughput is bounded by TypeSafe's 1,200 req/min (20 rows/sec) limit; horizontal scaling across worker nodes targets 1,500+ rows/sec cluster throughput.
• ~$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.
• Zero Rewriting: Emits clean records verbatim without rewriting or altering mathematical formulas.

---

Quickstart

Install (pick your OS)

OS	Rust	Python
Linux	cargo install jev-curate or gh release download tar.gz	pip install jev-curate (manylinux wheel)
macOS	cargo install or tar.gz (x86_64 + arm64)	pip install (universal2 wheel)
Windows	cargo install or zip	pip install (win wheel)

# Rust (any OS)
cargo install jev-curate
# or download prebuilt from GitHub Releases

# Python
pip install jev-curate  # no Rust toolchain needed (wheels on PyPI)
# from source: maturin develop  # python feature auto-enabled

# Set your TypeSafe AI key
export TYPESAFE_API_KEY="your-api-key"  # or set TYPESAFE_API_KEY in env

# Filter a Parquet dataset using the math reasoning preset:
jev-curate filter train.parquet \
  --preset reasoning-math \
  --model jev-latest \
  --out ./output/ \
  --concurrency 32
# YAML rubric: --preset rubric.yaml
# Parquet out: --format parquet (opt-in, keeps schema; else jsonl)

Python API

PyO3 bindings (build from source with the python cargo feature):

git clone https://github.com/AkashPriyadarshii/jev-curate
cd jev-curate
maturin develop   # python feature auto-enabled via pyproject.toml

from jev_curate import PyJevCurator

curator = PyJevCurator(api_key="your-api-key", preset="reasoning-math")
# single row
verdict = curator.filter_text("Let x=5... therefore...")  # -> {passed, rejection_reasons, scores, nouls}
# whole file -> ./curated/clean.jsonl + rejected.jsonl + errors.jsonl
stats = curator.filter_file("train.parquet", out="./curated")  # -> {passed, rejected, errors, total}

---

CLI Reference

jev-curate filter [OPTIONS] <INPUT_PATH>

Flag	Default	Description
<INPUT_PATH>	Required	Path to input .parquet or .jsonl file.
-p, --preset	reasoning-math	Pre-built rubric (reasoning-math, anti-sycophancy, code-correctness).
-o, --out	./curated/	Destination folder for clean.jsonl and rejected.jsonl.
-c, --concurrency	32	Worker concurrency (adaptive token bucket prevents 429 rate limits).
--dry-run	false	Host pre-filtering only, no API calls (no TYPESAFE_API_KEY needed).
--model	jev-latest	Jev model (or JEV_MODEL env, pin jev-1.13.0 for repro).
--format	jsonl	Output format: jsonl or parquet (opt-in, preserves schema).
--endpoint	None	Custom API endpoint URL for offline mock testing (or set TYPESAFE_ENDPOINT).

---

Presets

Preset	Primitives Evaluated	Target Problem Solved
reasoning-math	has_circular_reasoning (Noul) reasoning_depth (Score 0-4, keep 2.0 or higher)	Drops ungrounded math derivations and repetitive circular proofs.
anti-sycophancy	is_sycophantic (Noul) has_ai_disclaimer (Noul)	Eliminates "As an AI...", ungrounded flattery, and conversational filler.
code-correctness	has_stub_placeholders (Noul) code_quality (Score 0-4, keep 2.0 or higher)	Drops incomplete code blocks and unrunnable pseudo-code mocks.

---

Build & Test

# Rust core
cargo build --release
cargo test

# Python bindings (via maturin)
maturin develop
pytest

All tests run against an in-process mock server with zero live API credits in CI.

---

Architecture

jev-curate/
├── Cargo.toml                 # Rust core manifest (arrow, parquet, tokio, pyo3, clap, reqwest, serde)
├── pyproject.toml             # Maturin Python package manifest
├── src/
│   ├── lib.rs                 # PyO3 module bindings & crate entry
│   ├── main.rs                # Standalone CLI binary entrypoint
│   ├── client.rs              # TypeSafe AI HTTP client (speculative fan-out)
│   ├── filter.rs              # Host-side sanity pruning & Jev pipeline
│   ├── parquet_io.rs          # Batch-decoded Parquet reader + buffered JSONL writer (clean/rejected/errors)
│   ├── rate_limiter.rs        # Request rate limiter with auto 429 backoff
│   └── presets.rs             # Pre-built post-training evaluation rubrics
└── tests/
    └── mock_test.rs           # In-process mock tests via typesafe-rs-mock (100% offline)

---

Non-Goals

1. Not a Generative Re-writer: jev-curate never paraphrases or re-generates text. Data is kept 100% verbatim.
2. Not a Heavy Local Vector DB: No embeddings, no vector indices, zero PyTorch/CUDA runtime requirements.
3. Not a Generic Web Scraper: Tailored strictly for structured datasets (Parquet, Arrow, JSONL).

---

Mandatory Ecosystem, Author & Social Directory

Ecosystem

• jev-seo
• jev-superpowers
• jev-curate
• jev-git
• tdlib-android
• kharcha

Author

• Akash Priyadarshi (Patna, Bihar, India)
• Akash Priyadarshi, self-taught developer & open-source builder from Patna, Bihar, India. Rust systems tooling, AI agents, Android internals.
• GitHub • Portfolio • Project page • LinkedIn • Resume

Social

• X / Twitter • Threads • Instagram • Reddit

---

Built with high-performance Rust for the TypeSafe AI System One (Jev) ecosystem.

Keywords: TypeSafe AI, Jev, api.typesafe.ai, System One, Choice, Score, Noul, dataset curation, synthetic data filtering, pretraining datasets, Parquet streaming, arrow, rust.