JEV DataOps
Workbench where Jev sorts training records into keep, review or reject, then trains LoRA adapters and evaluates them.
Overview
JEV DataOps turns a dataset into an inspectable fine-tuning experiment. You upload data, Jev screens each record into keep, review or reject with a reason, the tool splits kept records into training, validation and test sets, trains LoRA adapters for a chosen Hugging Face model, and compares loss on held-out data. Every run saves screening decisions, splits and evaluation reports.
It is for people building small domain datasets who want a traceable pipeline. It offers browser, CLI, Python and HTTP interfaces, general, finance and code screening rubrics that can be customised, and exports for SFT or verl GRPO/PPO training. A public demo runs entirely in the browser with no key or GPU, but it uses local rules and a small statistical model instead of calling Jev, so its numbers show the workflow rather than Jev's screening quality.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Should this training record be kept, reviewed or rejected? | keep · review · reject |
How it works
- Upload JSONL or CSV data
- Jev screens each record against a rubric into keep, review or reject
- Split kept records and train LoRA adapters
- Evaluate on held-out data and save all reports
Strengths
- Per-record audit logs and reasons
- Multiple interfaces and export formats
- Customisable domain rubrics
Caveats
- Public demo does not call Jev or train a model
- Training requires GPU resources locally
Getting started
Try the browser demo at https://jev-dataops.vercel.app/demo/ or follow the README quickstart to run locally.At a glance
- Type
- Model
- Status
- beta
- Language
- Python
- Runtime
- Web workbench, CLI and HTTP API
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
An open-source JEV-powered workbench for streaming data selection, quality evaluation, automatic LoRA training and held-out model evaluation.
logicrw/awesome-jev-projectsAn open-source JEV-powered workbench for streaming data selection, quality evaluation, automatic LoRA training and held-out model evaluation.
hellogumbo/awesome-jevTraceable pipeline for training data: upload, screen with Jev, evaluate, fine-tune your own model, then evaluate the result, through a browser workbench or a CLI.
MrJev/awesome-jev
Frequently asked questions
What does Jev decide?
Whether each record should be kept, reviewed or rejected for training, with a reason.
Does the demo use Jev?
No. The browser demo uses local rules and a small statistical model to show the workflow.
Can I use my own training stack?
Yes, it exports SFT or verl GRPO/PPO bundles.
Related projects
dbt package whose macros run Jev Choice, Noul and Score over SQL values in DuckDB and ClickHouse models.
ProjectbetaSQL (Jinja)DuckDB extension that asks Jev about every row in SQL and returns answers as ENUM, DOUBLE or STRUCT columns.
PluginbetaC++Polars library for parallel LLM inference that also asks Jev per-row Noul, Choice and Score questions through a native Rust layer.
LibrarybetaPythonTabular classifier that learns which slices of labeled rows and columns to send Jev or LAYA, stopping when it has enough.
ProjectresearchPythonSutro's active-learning CLI that finds rows Jev is least sure about, asks you to label them, and uses GEPA to improve the definition.
CLIbetaPythonCLI that scans tabular, structured and text files or databases and asks Jev which fields look anomalous, building a review queue.
CLIbetaPython