User repository on GitHub
Open, local, zero-cost reimplementation of the Jev decision layer for job postings
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Practices & Patterns
Openjev — DECRUX9812
Open local classifier that reproduces a hosted Jev job-posting decision with a frozen CPU encoder plus small per-question heads, answering the same posting questions offline at zero marginal cost.
- Stars
- 3
- Language
- Python
- License
- MIT
- Published by
- Community
- Author
- DECRUX9812
- Last verified
- Oct 2026
This is one of several Jev listings for the same project. The canonical entry is OpenJev — razorback16.
About this project
An open, local, zero-cost reimplementation of the Jev decision layer for job postings. Jev is a closed, hosted model that answers seven typed questions about a job posting so a firm can tell "is this an in-house IT requisition, or a small business that needs a website?" That distinction is the whole ballgame: it is the difference between a posting that is a recruiting signal and one that is a sales lead. openjev answers the same questions, on your own CPU, with no API key, no network, and no per-call cost.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- Python
Highlights
- Frozen encoder plus small heads keeps the trained part to a few hundred KB.
- Same minimal document format as the closed model for a fair comparison.
- CLI and Python API return bucket, probabilities, sub-answers and lead flag.
- Ledger documents baselines, retrieval blends and synthetic-data results.
Watch out
The service_lead class has one real labelled example in the held-out set so it is not reliably detectable, the stratified gold set over-samples rare classes, and the training postings are not redistributed. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
Yes. The digest reports the shipped model close to hosted Jev on the same held-out postings with very high bucket agreement across the full production corpus; the eval script prints gold accuracy, recall and agreement separately.
A frozen quantised ONNX encoder plus tiny MLPs trained on a few thousand labelled postings in minutes on laptop CPU. The encoder is never fine-tuned, so reruns are byte-identical.
It sees a minimal title, employer and pay string matching what the closed model gets, deliberately without the legacy category column being replaced. That keeps the comparison fair.
No. Synthetic mixtures helped rare classes alone but hurt the real-postings boundary in every mixture tried, so the shipped weights exclude them; the ledger records the negative result.
Telling a recruiting signal from a sales lead. The seven typed questions ask whether a posting is an in-house IT requisition or a small business that needs a website. A Python caller gets a bucket, bucket probabilities, the five boolean sub-answers plus fit, and an actionable_service_lead flag; the CLI mirrors that at python -m openjev.cli with --employer and --pay.
service_lead has exactly one real labelled example in the entire corpus, and it sits in the held-out set, so that class is not learnable from real data alone and the project does not claim to detect it reliably. The gold set is also stratified, deliberately over-sampling rare classes, so its accuracy is not what you would see on a raw production stream.
Yes. python -m openjev.eval --gold-data <dir> --corpus <file> --weights weights, pointed at a directory holding leads.sample.jsonl and leads.gold.json, prints gold accuracy, per-class recall and, separately, agreement with hosted Jev across the full production corpus. Every number in docs/approaches-tried.md is on the same 70 held-out postings, and the training corpus itself is not redistributed.
More Jev resources like this
24GitHub stars
Unofficial laptop study of parallel constrained decoding: prefill once, share the KV cache across fields, and decide all schema fields in one batched forward pass without writing JSON.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
rorshopping/jev-on-a-laptopPython
rorshopping/jev-on-a-laptopPython22GitHub stars
Offline evaluator that checks whether classifier confidences match reality and converts mistake costs into a deployable threshold, emitting an HTML report plus machine-readable threshold files.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
rlaope/jevalPython
rlaope/jevalPython16GitHub stars
Vendor-neutral harness that tests whether Jev reranking improves a small RAG stack, using frozen hybrid BM25 plus BGE-M3 candidates and isolating retrieval, rerank, and generation stages.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
erendikmenn/jev-rag-benchmarkPython
erendikmenn/jev-rag-benchmarkPython37GitHub stars
Minimal Python recipe that turns any causal LLM into a Jev-style classifier by prefilling choice_index and constraining sampling to legal indexes in one forward pass.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
sgoedecke/system-onePython
sgoedecke/system-onePython39GitHub stars
Open replica of the typed decision API on openly available base models, answering choice, score, and true-probability questions in one forward pass with no decoding.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
mithalouni/system-one-openPython
mithalouni/system-one-openPython941GitHub stars
Type-safe generation layer over existing autoregressive models via JSON Schema on SGLang, returning constrained string, numeric, boolean, and enum answers with optional thinking.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
zmtomorrow/TypeARPython
zmtomorrow/TypeARPythonFrom the Jev guides
Original write-ups on what this entry is for.
Most Jev accuracy problems are state problems. What belongs in a state, what is a pre-made judgment to leave out, and why naming the slice you ask about matters more than the wording.
Jev Practices & Patterns#state#design#primitives
Read article
A Noul returns one number and no separate confidence. How to read the value, pin its boundary with criteria, split compound questions, and set a per-action threshold that matches the cost of being wrong.
Jev Practices & Patterns#noul#primitives#confidence
Read article
A Score is a weighted position, not a ruler. Why the number can land between levels, why levels are judged separately, and how to combine several atomic scores in code instead of one crowded rubric.
Jev Practices & Patterns#score#primitives#scoring
Read article
From the Jev community
Posts from builders shipping with Jev right now.