FRIDA Decisions: When Judges Need No Generation

Sberbank made publicly available FRIDA Decisions.

It is a CIS-language model that selects an option from a list, assigns a score on a scale, answers “yes” or “no,” and ranks candidates in a single encoder pass.

It does not generate text, and tens of milliseconds are enough for it on an RTX 5060 Ti.

It scores 0.893 on the razvilka benchmark.

The commercial TypeSafe Jev API scores 0.897, while the Qwen3.5-35B-A3B LLM, which is 40 times larger, delivers comparable quality.

My hypothesis is that in enterprise AI scenarios, classifier tasks account for a significant share of calls to a large generative model, and are paid for as generation.

In many LLM implementations, the call looks the same: the model receives text and a list of labels, then returns a label.

This is how support request routing, moderation, data labeling, checking another model's answers (the “judge” role), and ranking snippets for RAG work.

For such a call, the company pays for generation tokens, waits hundreds of milliseconds or seconds, and receives an answer in an unpredictable format.

At a volume of tens of thousands of requests per day, this becomes a separate budget item and a gateway queue.

If the judge is called for every assistant response, its latency is added to the latency of the entire chain.

Tens of milliseconds on a single consumer-grade GPU change the economics of such a solution.

The comparison on razvilka gives three data points: the open FRIDA Decisions encoder scores 0.893, the commercial TypeSafe Jev API scores 0.897, and Qwen3.5-35B-A3B (an MoE model with 35 billion parameters) reaches a comparable level.

There are four thousandths between the first and second scores.

For routing and moderation, a confidence threshold outweighs this difference. This result has limits.

The benchmark measures a specific set of tasks, while your requests, jargon, and complex topics may fall outside its distribution.

Therefore, razvilka provides a reason to test the model on your own labeled data; it does not replace such a test.

We would start with a sample of a couple thousand historical decisions for which the correct answer is already known.