Kev

Kev

Scores typed questions about a document with small Kev decision models, returning a probability for every answer option in one pass and no text. It works in the browser, from Python or as an MCP tool.

Made by jaredpalmer on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: Kev is a set of compact decision models from jaredpalmer, built on top of Qwen3.5, and this Gradio Space lets you try two of them. You give it one document and some typed questions. For each question it scores every answer option with a probability, using a single pass over the document. It does not write any text.

Who it suits: Developers who want odds for classification-style questions about a text, such as routing a message, and people who compare small decision models.

Industry | Who uses it | How they use it | Example
Customer support | Support engineers | Route a message to the right team by probability | Asking which team should handle a late, wrong-sized order
Machine learning | Model evaluators | Compare the 4B and 0.8B models on the same request | Choosing Both in the Space
Software development | Agent builders | Call the decide endpoint from code or as an MCP tool | Reading the answers from a Python client

Skip it if you need written answers or summaries, since Kev returns probabilities and no text.

Where to use it: Kev works in these places.

• A web browser: pick a model, or both, and press Decide
• Python: call the decide endpoint with the gradio_client package
• An MCP client: the decide endpoint is also offered as an MCP tool
• Python 3.12 on Windows, macOS or Linux: run a copy of the Space

Use case | Who | What they get back
Route a message | Support teams | A probability for each team option and the most likely one
Check a yes or no | Developers | The probability that a statement is true
Rate on a scale | Reviewers | An expected level and a probability for each level

How to set it up: Online use needs only a browser. A local copy or the Python client needs a few things first.

• A web browser for the online Space
• Python 3.12 on Windows, macOS or Linux for a local copy
• The Python packages listed in requirements.txt
• The gradio_client package, if you call the endpoint from Python

1. Install the Hugging Face hub library with pip.
2. Download the Space's files into a folder called kev.
3. Move into that folder.
4. Install Gradio at the version the Space uses.
5. Install the packages listed in requirements.txt.
6. Start the app with Python.

pip install -U huggingface_hub
hf download jaredpalmer/kev --repo-type space --local-dir kev
cd kev
pip install gradio==6.28.0
pip install -r requirements.txt
python app.py

How to use it: Online, pick Kev-4B, Kev-0.8B or Both, paste a document and your typed questions, then press Decide. Three options sit next to the button. From code, call the decide endpoint with the Python client below.

from gradio_client import Client
c = Client("jaredpalmer/kev")
rendered, response, report = c.predict(
    "Shoes arrived two weeks late and in the wrong size.",
    '{"department": {"type": "choice", "instructions": "Which team should handle this?", "criteria": {"returns": null, "shipping": null, "billing": null}}}',
    "Kev-4B", False, False, False, 4,
    api_name="/decide",
)
print(response["answers"])

Tool | What it does
choice | Picks among named options and returns the top option, a probability for each and a confidence
noul | A yes or no question that returns the probability of true
score | Rates against an ordered list of level descriptions and returns the expected level and a probability per level
Calibrated probabilities | Applies one temperature fitted on development data, which keeps the top choice and lowers out-of-domain error
date_facts | Adds the day count between every pair of dates in the document before the model reads it
Option-order stability | Re-runs the first choice question with shuffled options and reports whether the top choice changes
decide | The single entry point for code and MCP clients, returning the answers for your questions

Example: You paste a note saying a pair of shoes came late and in the wrong size, and ask which team should pick it up. The options are returns, shipping and billing. Kev returns a probability for each team and names the most likely one.

For newcomers and experts: If you are new, think of Kev as a fast way to ask a multiple-choice question about a text and get odds for each answer instead of a written reply. Try one of the built-in examples first, such as Return window.

If you know the field, a Kev checkpoint is a LoRA adapter folded into a Qwen3.5 base model. The vocabulary head is discarded, and a pointer head matches each question's decide token to its option spans. Questions run in separate rows that all continue from the shared document, so a secret placed in one question stays hidden from the others. The trade-off is calibration outside the training domain.

You supply a document and typed questions
One pass reads the shared document
Each question runs in its own isolated row
A pointer head scores the options
You get a probability for every option

Good to know: Kev gives probabilities, not text, and the raw numbers are imperfect outside the training domain.

• Calibration: outside the domain it was trained on, the numbers help but are not exact, and the 0.8B model does clearly worse there
• Gaps: a question unlike anything in training may get an unreliable answer, so test it on your own data first
• Dates: Kev cannot subtract dates by itself, and the date_facts option adds the day counts
• Version: Gradio 6.28.0 and Python 3.12
• Models and data: two Kev checkpoints built on Qwen3.5 base models, plus one named dataset of test suites
• Size: 13 files in the repository
• License: apache-2, and the Qwen base models are Apache-2.0 too
• Cost to run: none named, and the Space uses zero-a10g hardware at Hugging Face
• Data it reaches: the document and questions you send, and the models it names
• Not for: generating text, since Kev returns probabilities only

Title: Kev-4B out-of-domain calibration error, lower is better
Raw probabilities: 0.12 ECE
Calibrated: 0.05 ECE

On Hugging Face, 15 November 2026

• Likes: 96
• Built with: gradio
• License: apache-2.0

Links

• Space: https://huggingface.co/spaces/jaredpalmer/kev
• Files: https://huggingface.co/spaces/jaredpalmer/kev/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.