Benchmarks & research

FRIDA-Decisions

SberAI's open FRIDA-Decisions scores Russian-first inputs locally to return fixed answers and rankings rather than generated prose.

Bookmark Keep this in your collection.

Leave a noteWhat would you try with this?

Only you can see your notes.

Source screenshot of FRIDA-Decisions — SOURCE SCREENSHOTFull screenshot ↗

What it does

The MIT-licensed encoder has about 823M parameters and local serving options. Russian/English comparisons and hardware timings are the maker's measurements. Training covered text up to 512 tokens, but the documented default cuts at 384. Ranking cutoffs need task-specific testing. This Jev competitor does not use TypeSafe weights.

How you can use it

Start with the supplied local examples, a Russian-language task and clearly described options. Check how much input will be scored. Choose ranking cutoffs using cases with known answers; the default length and score settings do not show that longer documents or every ranking task will work.