tab-jev
Combines Jev judgments on row text with a tabular foundation model like TabPFN to predict outcomes from few labels.
Overview
tab-jev is a Python library for classification on data that mixes text and tables. A jev backend reads each row's text and answers a Noul on the target plus Score rubrics, in Jev's request and response shape; every option's probability becomes a new numeric column. A tab backend, any scikit-learn style classifier such as TabPFN, then learns from the labeled rows in context using those columns and the original table fields. Jev can answer on day one with no labels, and tab improves as labels arrive.
It targets problems with a few dozen to a few thousand labels where both text and table carry signal and decisions act on the probability, such as churn, fraud, claims screening, lead scoring and ticket escalation. On a Kickstarter funding dataset the README reports 0.745 AUC at 256 labels versus 0.682 for calibrated Jev alone. A local option pairs the open Kev model with TabICL so no data leaves the machine. The project is pre-release and hosts no models itself.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Will this Kickstarter project reach its funding goal? | — |
| Score | Rubric questions about each row's text, used as features | — |
How it works
- Define the target as a Jev question and optional Score rubrics.
- Jev answers each row's text; probabilities become numeric columns.
- A tabular foundation model fits on labeled rows plus these columns.
- Predict probabilities for new rows and act on thresholds.
Strengths
- Useful predictions with very few labels.
- Pluggable backends, including a fully local Kev plus TabICL pipeline.
- Outputs stay in Jev answer format.
Caveats
- Pre-release; API will change.
- Hosted backends send rows to TypeSafe and TabPFN services.
- Smaller local models give lower scores.
Getting started
Use jev_then_tab(target, JevHTTP(), TabPFNClassifier(), rubrics=..., jev_columns=[...]); install the local extra with pip install "tab-jev[local]".At a glance
- Type
- Library
- Status
- prototype
- Language
- Python
- Runtime
- library
- Has code
- yes
Use cases and integrations
How curated lists describe it
Tabular prediction: asks Jev a `Noul` on the target plus `Score` rubrics about each row's text, turns every option's probability into a column next to the row's numeric fields, and lets a tabular foundation model such as TabPFN learn from the labeled rows in context, reaching 0.745 AUC at 256 labels on Kickstarter funding against 0.682 for Jev alone with calibration.
yibie/awesome-jev
Frequently asked questions
What does Jev contribute?
It turns each row's text into probabilities for the target and rubric questions, which the tabular model uses as features.
When should I use Jev alone instead?
The README suggests Jev alone when you have no labels, and Jev plus calibration when the label is a judgment of the text itself.
Can it run without sending data out?
Yes, the local option uses Kev for text and TabICL for learning, both on your machine.
Related projects
Self-hostable Reddit buyer-intent finder that has Jev judge every post and comment it reads, with a written reason and cost per lead.
ProjectproductionTypeScriptScores HeyReach LinkedIn leads and first-message variants with Jev, picks a message per lead and writes draft campaigns back.
ProjectbetaPythonFloating live-call coach that transcribes locally and asks Jev after each sentence what to do next, shown with a confidence gauge.
ProjectprototypeJavaScriptTypeScript CLI and REST API that asks Jev to rate an occupation's AI layoff risk, trajectory and need for human accountability.
CLIprototypeTypeScriptStartup-idea evaluator where Jev scores eight indie-hacker questions and code turns the weighted average into KILL, FIX or SHIP.
AppprototypeTypeScriptLocal China A-share technical analysis workbench where Jev assesses eight indicator dimensions and returns Buy, Hold or Sell with evidence.
ProjectbetaPython, TypeScript