SberAI releases FRIDA-Decisions, a Russian and English encoder for structured decisions
A 0.8B MIT-licensed encoder that picks, rates, answers yes/no or ranks in one pass. The card reports 0.893 on 735 razvilka items.
SberAI (listed on the card as УЭСМО, SberAI) published FRIDA-Decisions on Hugging Face on 2 October 2026, by repository creation date. The card describes a Russian-language encoder model for structured decisions on text. It picks one option from a set, places text on an ordinal scale, answers yes or no, or ranks candidates, all in a single encoder pass with no text generation.
It has about 0.8B parameters, is built on ai-forever/FRIDA, a T5 encoder, and is licensed MIT. Training data is 71% Russian and 29% English. Inputs are trained up to 512 tokens, and longer texts are cut by default.
On the razvilka benchmark of 735 items the card reports 0.893, with an int8 ONNX CPU build at 0.891. It lists TypeSafe Jev at 0.897 and says the difference is not statistically significant (paired McNemar p = 0.84). It calls FRIDA-Decisions the highest of the open models it tested on razvilka. Speed is about 28 to 34 ms per request on an RTX 5060 Ti.
The benchmark is a third-party dataset linked from the card. The comparison is the maker's.
A small encoder, an open licence, and a language the English-first models skip.
Sources
Editor's first pass (AI-assisted)
- Real 4: MIT weights, GitHub repo and demo Space are linked.
- Useful 3.5: Fills a Russian-language gap with a 0.8B model that runs on a CPU build.
- Notable 3: Shows decision models spreading beyond English.
- Novel 3.5: An encoder rather than a decoder base, aimed at one language pair.
Comments