Youssef

MK-SH-V

MK-SH-V: a 260M visual decision model. Calibrated probabilities, a choice, and a refusal below a validated gate, over document pages and screenshots.

Code on GitHub Try the demoModel

accuracy on held-out real screenshots; a model 15× larger scores 0.731 zero-shot
0.839

The released checkpoint on the held-out real-screenshot test. The refusal gate was fitted on a separate calibration split, by a rule frozen in advance.

Language
Python
Licence
Apache-2.0
Last push
31 Oct 2026
Stars
5

What it is

A small open visual decision model. Give it a document page or a screenshot, a question and the options; it returns calibrated probabilities and a choice, or it refuses when its confidence sits below a gate whose risk was validated on held-out data. The visual sibling of MK-SH-1.

How it works

Three real cases, run through the released weights and the shipped policy:

The receipt’s total is 302.016 IDR. The E41 release refused the second question at 0.507; this one reads the total and answers it. The third row is where it stops: the right range is “above 200.000”, and the model picked “below 100.000” with a confidence that clears the gate. On real receipts about a quarter of the answers the gate lets through are wrong (see the gate, held out); the gate’s risk budget is validated on the synthetic test only.

How a decision is made: one encoder pass reads the image and the text together, each option is marked inline and scored by a shared head, and the softmax runs over the valid options only. The temperature T only rescales confidence, so it never changes a choice. The gate tau was fitted on a calibration split (E48, by the E42 protocol) as the smallest threshold whose Clopper-Pearson 95 % upper bound on selective risk stays within a 5 % budget. Images whose longest side is under 1,024 px are first enlarged to it (Lanczos), exactly as in training and calibration (E46). It is not a chat model and generates no text.

Results

0.839 τ = 0.55 98.3 %
accuracy on real screenshots, held out; a ~15× larger model run zero-shot scores 0.731 the refusal gate, fitted on a calibration split by a rule frozen in advance answered on the held-out synthetic test, with a 95 % Clopper-Pearson risk bound of 3.2 %

Try it

pip install "mksh-v @ git+https://github.com/Gjusev/mk-sh-v"