FRIDA-Decisions

FRIDA-Decisions

Makes structured decisions over Russian text in one pass: pick an option, rate on a scale, answer yes or no, or rank candidates. Made for developers who classify and route Russian messages.

Made by ai-forever 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: FRIDA-Decisions reads a Russian text and answers questions about it by choosing among options you declare. It can pick one of several labels, place the text on a scale, answer yes or no, or rank candidates. It never writes free text, so every answer is one of your options, with a confidence.

Who it suits: It suits developers who sort, triage or route Russian messages and want a fast, predictable classifier whose labels they can change without retraining.

Industry | Who uses it | How they use it | Example
Customer support | Helpdesk tool builders | Sort incoming Russian messages by topic and urgency | Label a message about a missing code as a login issue
Content moderation | Platform developers | Flag spam in user messages | Answer yes or no on whether a message is spam
Software | App developers | Match a user request to one of many intents | Check a catalog of 243 intents in one pass
Research | NLP researchers | Test decision accuracy on a Russian benchmark | Run the model on the razvilka set

If you need text generation, or your texts are in other languages, you can skip it.

Where to use it: It runs locally in Python, on a consumer GPU, on a CPU through an ONNX build, and on Apple Silicon through an MLX build.

• Python with PyTorch on an NVIDIA GPU
• CPU-only machines, using the ONNX build
• Apple Silicon Macs, using the MLX build
• A vLLM server on Linux with a GPU, for many users at once
• A Colab notebook the model card provides

Use case | Who | What they get back
Topic sorting | Support teams | One declared label per message, with its confidence
Urgency rating | Triage tools | A position on the scale you define
Spam check | Moderators | A yes or no answer with its confidence

How to set it up: You need Python with pip, and the frida-decisions package, which the model card pins at version 0.4.0 with its torch extra. For CPU use without PyTorch, the card names an onnx extra instead.

1. Install the Hugging Face hub tools.
2. Download the model files.
3. Install the frida-decisions package from the card, with its torch or onnx extra.
4. Load the model in Python with the Judge class.

pip install -U huggingface_hub
hf download ai-forever/FRIDA-Decisions

How to use it: You build a request with the text as the state and a set of named questions. Each question has a type, instructions and criteria. The types are choice, score and noul, which means yes or no.

from frida_decisions import Judge

judge = Judge.from_pretrained("ai-forever/FRIDA-Decisions", device="cuda")

Pass the request to the judge and print the answers.

print(judge.judge(request)["answers"])

Tool | What it does
Judge | Runs the model in process, here on a CUDA GPU
OnnxJudge | Runs the int8 ONNX build on a CPU
vLLM server | Batches requests from many users on Linux with a GPU
hf download | Fetches the model files to your computer

Example: You send a Russian support message saying a confirmation code has not arrived for an hour. You attach three questions: its topic among login, payment and delivery, its urgency, and whether it is spam. Each answer comes back as one of your declared options, with a confidence.

For newcomers and experts: Think of it as a form filled in by a reader. You write the question and the allowed answers, and the model ticks one. A good first try is the support-message example with two or three topics.

It is built on the FRIDA T5 encoder with 823M parameters. Packing puts each question's options into a single sequence, so the text is read only once and follow-up questions cost little. The card reports 0.893 on the razvilka benchmark of 735 items, against 0.897 for a commercial API.

Write the text and your questions as JSON
The text is encoded once in one pass
All declared options are scored together
Each question returns one option and a confidence

Good to know: Packing all options into one sequence is much faster than scoring each option separately, according to the card.

Title: Time to answer a catalog of 243 intents
One packed sequence: 0.44 seconds
One sequence per option: 4.65 seconds

• Speed: on an RTX 5060 Ti the card measures 28 to 34 ms for each request of about 400 tokens.
• Memory: PyTorch peaks near 1.8 GiB over the whole razvilka run, with the CUDA context on top.
• ONNX on CPU scores 0.891 on razvilka and takes about 0.9 s for a 384-token request with 3 questions on 6 threads.
• License: MIT.
• Cost to run: none named. It runs on your own hardware.
• Data it reaches: the Russian text you pass in, handled locally unless you use a vLLM server.
• Not for: free-form generation. It returns only the options you declare.

On Hugging Face, 13 November 2026

• Task: text classification
• Library: frida-decisions
• Parameters: 823M
• Weights: about 1.6 GB in BF16; about 412 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 1,111
• Likes: 57
• License: mit
• Languages: ru, en
• Based on: ai-forever/FRIDA
• Tags: frida-decisions, onnx, safetensors, t5, decisions, routing, intent-classification, guardrails, zero-shot-classification, reranking, encoder, russian

Links

• Model: https://huggingface.co/ai-forever/FRIDA-Decisions
• Files: https://huggingface.co/ai-forever/FRIDA-Decisions/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download ai-forever/FRIDA-Decisions

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