Are the accuracy claims real for the model I would download?

Only the older 68M baseline has a published typed-decisions score beating Jev in the repo; the current 150M release reports ONNX parity and routing numbers instead, so compare them as different releases.

Can I let it auto-triage CI failures and block merges?

Poorly: the repo reports very low accuracy on its hand-labelled CI-failure set and says to keep that workflow in shadow mode with human review, never blocking or releasing on its own.

What breaks most often on first install of the local release?

Run git lfs pull after cloning so the ONNX graphs are real files, then load the Hub id or the local release directory; large model files arrive separately from the code.

How do I swap it in for Jev without rewriting my client?

Point an existing Jev or Laya client at the serve endpoint, or use Engine.load and predict in Python; every answer carries an abstain flag and per-option probabilities you can threshold.

Does the exported model need a GPU?

No. Engine.load takes an exported model directory and runs on the CPU only via ONNX, with Python 3.10+ required. Training is a separate story, since the notebook expects a GPU on Colab or Kaggle, but inference itself is CPU-bound.

What is it built on, and under what license?

Apache-2.0. It is a packed Ettin-150M encoder from JHU CLSP (MIT), streams pages from Common Crawl, and labels them from each page's own schema.org markup while the model never sees that markup. The /v1/systemone format follows Jev's public API.

What does it return, and does it abstain?

For each question it returns a probability for every answer and abstains when unsure, so only hard pages go to an LLM or a person. The question types are choice (one of the named options plus a probability each), noul (probability of true) and score (expected level plus a probability per level).

CommunityJev Agents & MCP

Anarkali

Compact open-weight typed-decision model with a local 150M ONNX release for CPU inference, speaking the Jev and Laya /v1/systemone request shape.

See highlights

Open github.com

Stars: 3
Language: Python
License: APACHE-2.0
Published by: Community
Author: ToufiqQureshi
Last verified: Oct 2026

ToufiqQureshi/anarkali

User repository on GitHub

View on GitHub

compact 150M typed-decision engine: calibrated probabilities, abstain flag, Jev-compatible /v1/systemone API, local ONNX on CPU. 0.3.0 (68M) beat Jev on typed-decisions.

Forks: 0
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

About this project

A small, fast decision model for web data work: scraping, crawling, and price, SEO and competitor monitoring. Anarkali sits in front of your scraper or your LLM pipeline and answers small typed questions about a web page: what kind of page is this? which text is the price? is it in stock? It runs on a CPU, returns a probability for every answer, and abstains when it is unsure, so only the hard pages go to an LLM or a person.

Quoted from the project's README. Read the full README.

Tagscommunity github

Works forJev for Classification

Used inJev for Software & DevTools

Highlights

Local 150M ONNX release with tokenizer and config for CPU inference, plus a Hub model id for one-line loading.
Speaks the /v1/systemone format so existing Jev and Laya clients work by pointing at a new URL.
Every answer carries an abstain flag with calibrated probabilities, and invalid schemas are rejected before inference.
Ships a 32-case hard routing suite plus a hand-labelled CI-failure set with reproduce scripts.

Watch out

The 150M release has not been scored on the public typed-decisions benchmark, and the repo says the CI-failure workflow is not production-ready. 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.

More Jev resources like this

wald-4b

Open-weight 4B decision model on Qwen3.5-4B-Base serving a Jev-compatible /v1/systemone API, returning per-option probabilities with adjustable thinking.