Organization repository on GitHub
Pluggable framework for predictions on mixed text and tabular data (early development)
- Forks
- 1
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Repos & SDKs
Tab Jev
A pre-release Python framework for predictions on mixed text and tabular data: Jev Noul and Score questions turn text columns into numeric features, then a tabular foundation model such as TabPFN learns from labeled rows in context. Offers hosted and fully local backends.
- Stars
- 4
- Language
- Python
- License
- APACHE-2.0
- Published by
- Community
- Last verified
- Oct 2026
About this project
A jev-like model + a tabular foundation model = an in-context learner for your text & tabular data. Predictions on data that mixes text and tables, with pluggable backends: jev backends turn text into typed judgments (a chosen option plus probabilities), in the request/response shape of TypeSafe's Jev API. tab backends are tabular foundation models that predict by in-context learning. Any scikit-learn style classifier works, such as TabPFNClassifier.
Quoted from the project's README. Read the full README.
Highlights
- On Kickstarter funding data it reached 0.745 AUC at 256 labeled rows, versus 0.682 for Jev text alone with calibration and 0.688 for the tabular model alone.
- Rubric Score questions about a text column become per-option probability columns next to the row's numeric fields.
- Targets use Jev's noul, choice and score question types, and predictions come back in Jev's answer format.
- The fully local path pairs open Kev 0.8B with TabICL, both with licenses allowing commercial use, for data that cannot leave the machine.
- Presets include jev_then_tab, tab_then_jev and parallel_blend, or steps can be composed into a custom Pipeline.
Watch out
Apache-2.0 pre-release whose API the README says will change. Best results need paid Jev and TabPFN services, and the package itself hosts no models.
Questions about this entry
Short answers drawn from the linked project or page.
The README recommends it when text and the table each carry part of the signal and you have dozens to a few thousand labeled rows and act on the probability. With no labels it suggests Jev alone, and with 10k+ labels a supervised model.
Yes. The kev_to_tabicl path runs the open Kev 0.8B reader and TabICL on your machine with nothing leaving it. Small models are weaker readers than the Jev API, so expect lower scores that the tabular model can learn to discount.
jev answers your target question and any rubric questions about the text with no labels at all, then every option's probability becomes a new numeric column beside your existing ones. The text columns drop out and the tabular model sees an all-numeric table. It then does in-context learning on the labeled rows with no training step, and with no labels or fewer than ten, the tabular model is skipped entirely and jev's own answer becomes the prediction.
One answer per row in Jev's own format, so code written against the Jev API reads it unchanged, for example a noul dict carrying P(funded). Targets are declared with Jev's noul, choice and score question types, as Question('noul', 'Will this project reach its funding goal?') or Question('score', 'How clear is the pitch?', ['vague', 'somewhat clear', 'clear', 'very clear']). Rubrics are optional; with none, jev only answers the target question.
On the Kickstarter funding dataset with 2,002 test projects, tab-jev reaches 0.745 AUC at 256 labeled rows against 0.682 for Jev text plus calibration and 0.688 for the tabular model alone, and 0.765 at 1,024 rows on a single label draw. At 64 rows the margin is thin: 0.684 against 0.682. The full method and numbers are in the demo directory.
jev_to_tabpfn is the recommended preset for the strongest reader and tabular model, installed with pip install tab-jev[tabpfn]. kev_to_tabicl ships with the package for when data cannot leave the machine or you want to try without API keys. There are also tab_then_jev and parallel_blend presets, and any can be composed into a custom Pipeline. The hosted backends send your rows to their services: jev_then_tab posts to the Jev API reading TYPESAFE_API_KEY and TabPFNClassifier posts reading TABPFN_TOKEN. The package hosts no models itself.
A rubric is a set of extra questions you ask jev about a text column, beyond the question you actually want answered, and it turns free text into numbers the tabular model can work with. The reasoning is that signals like how clear or how appealing a pitch is cannot be read off raw text by a table model. Write them as narrow, literal questions whose answer you can actually see in the text, such as how clear and concrete the pitch is on a four-level scale, and map a text column to named questions of the same three types as targets.
More Jev resources like this
21GitHub stars
Self-hosted Python SDK and System One-compatible API for runtime-defined Choice, Boolean and Score decisions on pinned open models, across Torch, vLLM, MLX, llama.cpp and WebGPU. Each result records probabilities, token usage, timing, model revision and prompt version, with committed row-level benchmarks.
Jev Repos & SDKs#community#sdk#self-hosted
Community
chengyongru/fastjevPython
chengyongru/fastjevPython2GitHub stars
Ruby gem that makes System One measurement a language primitive: psi prepares a state, judge/choose/score return calibrated distributions, and ? collapses them for control flow. Ships TypeSafe, protocol (/v1/systemone) and CUA providers behind one interface with its own spec suite.
Jev Repos & SDKs#community#ruby-gem#sdk
Community
innocentdiaz/s1_rubyRuby
innocentdiaz/s1_rubyRuby4.5kGitHub stars
Ruby-native AI framework whose RubyLLM::Judge class asks typed judgment questions: a probability that a statement holds, choices or scores, answered by the TypeSafe provider alongside 18 other hosted and local providers under one consistent API.
Jev Repos & SDKs#community#ruby#framework
Community
crmne/ruby_llmRuby
crmne/ruby_llmRuby77GitHub stars
An Apache-2.0 0.8B decision model that returns calibrated probabilities for yes/no, choice and score questions on CPU. Distributed as GGUF for llama.cpp and served behind a Jev-style /v1/systemone API, so existing Jev clients mostly need only a new base URL.
Jev Models & Reimplementations#community#gguf#local-inference
Community
kouhxp/gutsyHTML
kouhxp/gutsyHTML33kGitHub stars
Alibaba's multi-agent platform, which implements native TypeSafe Jev classification through JevClassifierModel for binary, choice and score routing across agent pipelines. Python 3.11+, Apache-2.0, with a FastAPI agent service and web UI.
Jev Integrations & Adapters#community#multi-agent#framework
Community
agentscope-ai/agentscopePython
agentscope-ai/agentscopePython882GitHub stars
A Node and TypeScript package that runs Convai's Laya checkpoint through ONNX Runtime, with the same system_one shape as the Python reference.
Jev Repos & SDKs#community#typescript#open-models
Community
receptron/layaTypeScript
receptron/layaTypeScriptFrom the Jev guides
Original write-ups on what this entry is for.
Closed weights have not stopped the ecosystem from rebuilding the shape in the open. API-compatible servers, retrained heads on small models, and inference-time tricks that skip generation entirely.
Jev Repos & SDKs#open-models#research#local
Read article
From the Jev community
Posts from builders shipping with Jev right now.