Benchmarks & research
FRIDA-Decisions is SberAI's Russian-first open model for fixed answers and ranking, with a local encoder that scores inputs instead of generating prose.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
The MIT-licensed encoder has about 823M parameters and local serving options. Its Russian/English comparisons and hardware timings are maker measurements. It was trained on text up to 512 tokens, while the documented default cuts at 384. Ranking cutoffs need testing for the task. It is a competitor to Jev, not TypeSafe weights.
Try the supplied local examples with a Russian-language task and clearly described options. Check the input length before scoring. For ranking, use examples whose answers you know to choose the cutoff. The default limit and score settings do not establish that every longer document or ranking task will work.