Ru-Jev

Ru-Jev

Scores a Russian-first decision from a context and a list of options, giving a probability for yes/no, choice or rating questions. Runs on a plain CPU, for teams that keep data in-house.

Made by besktech on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Ru-Jev is a Russian-first model that makes typed decisions instead of writing text. You give it a context and a list of options, and it scores every option with a probability and picks the likeliest. It handles yes/no questions, a choice from options and ratings on an ordinal scale.

Who it suits: It suits developers and data teams who need to route or classify Russian text, want a probability they can read, and work on a machine without a GPU or on a closed network.

Title: Who uses it
Industry | Who uses it | How they use it | Example
Software | Backend developers | Route incoming messages to a category | Pick weather from a list of 18 categories
Customer support | Support tooling teams | Sort Russian-language requests into fixed queues | Choose the right queue for a message
Closed networks | Platform engineers | Run the model on CPU inside a private perimeter | Score options with no outside service
Research | NLP researchers | Check how accuracy changes when a label is renamed | Compare a Russian word with a jargon wording

Anyone who needs free-form text generation can skip it, as can teams who work in languages other than Russian and English.

Where to use it:

• Python with the transformers library
• Windows, macOS or Linux with Python installed
• A CPU-only machine, no GPU needed
• A closed internal network

Title: Top jobs
Use case | Who | What they get back
Routing a message | Developers | A category from a fixed list, with its probability
Yes or no questions | Analysts | A probability for yes and one for no
Rating on a scale | Product teams | A probability for each level of the scale

How to set it up:

• Python on your computer
• The huggingface_hub package
• About 1.2 GB of disk for the full-precision weights

1. Install the Hugging Face hub package.
2. Download the model files to your computer.
3. Open the model card and copy its loading snippet. Do not use the stock zero-shot pipeline.

pip install -U huggingface_hub
hf download besktech/ru-jev

How to use it: Load the model with the snippet from its card, pass a context and your list of options, and read the probabilities it returns. Earlier releases sit on branches, and you pick one at load time with this argument.

revision="v1.1"

Example: You pass a short Russian message about tomorrow's rain together with a list of categories that includes weather. Back comes a probability for each category, with weather on top. On its own example the card reports 0.95 for the right option, where the stock zero-shot pipeline gave at most 0.44.

For newcomers and experts: Think of it as a sorter that reads a sentence and a list of choices, then tells you how sure it is about each one. The first small thing to try is a yes or no question in Russian, loaded the card's way. The stock pipeline returns random numbers, so avoid it.

Under the hood it is an mmBERT-base encoder with 307M parameters and a separate scoring head stored in head.safetensors. Stock pipelines load through AutoModelForSequenceClassification, which quietly starts a fresh head. The head begins from random weights and trains together with the backbone on open datasets with human labels. Version 1.3 trains each fixed option set in several wordings, so the model reads the option text instead of memorising it.

Context and options go in
The mmBERT-base encoder reads both
A separate head scores each option
Probabilities come out, top option picked

Good to know:

• Limits: The inference widget and the stock pipelines do not work with it, so load it the card's way.
• Weak spots: Opaque codes and short abbreviations as labels score poorly, and accuracy falls as the candidate list grows.
• License: MIT.
• Cost to run: none named, and it runs on CPU.
• Data it reaches: it downloads its files from Hugging Face, and the source names nothing else.
• Not for: generating text. It is not affiliated with TypeSafe AI.

Title: Accuracy on jargon labels with 5 candidates
v1.1: 0.633 accuracy
v1.3: 0.855 accuracy

On Hugging Face, 14 November 2026

• Task: text classification
• Library: transformers
• Parameters: 307M
• Weights: about 1.2 GB in F32; about 153 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 403
• Likes: 2
• License: mit
• Languages: ru, en
• Based on: jhu-clsp/mmBERT-base
• Tags: transformers, safetensors, modernbert, feature-extraction, text-classification, zero-shot-classification, typed-decision, russian, routing, intent-classification, calibrated, calibration

Links

• Model: https://huggingface.co/besktech/ru-jev
• Files: https://huggingface.co/besktech/ru-jev/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download besktech/ru-jev

Then follow the usage in its model card. Nothing in this kit runs by itself.