FRIDA-Decisions
Quadrant scores
See the full quadrantScored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.
- Maturity5.8/10
MIT weights (bf16 + int8 ONNX for CPU), a detailed card, GitHub package, Colab notebooks and a demo Space. Serves through its own vLLM pooling plugin (POST /pooling), not /v1/systemone, so the TypeSafe SDK does not point at it. pip install from git only. SberAI team, no SLA.
- Capability6.8/10
Choice / Score / Noul (plus ranking) from a T5 encoder with a scalar head. Returns a distribution per question, but no calibration claim, temperature or ECE. 28–34 ms in process on an RTX 5060 Ti, no p99 (author). razvilka 0.893 vs Jev 0.897 on a benchmark the same team built.
- Adoption4/100
Day one: 207 Hugging Face downloads and 15 likes, 3 GitHub stars, a demo Space and a couple of Russian-language mentions on X. No derivatives, integrations or production use reported.
Vendor claims
- razvilka (735 Russian items, 15 tasks): 0.893 vs TypeSafe Jev 0.897 on the same items (paired McNemar p = 0.84); highest among open models the team ran[Vendor claim — not independently verified]
- 28.2 ms (1 question) / 34.0 ms (3 questions) per request on an RTX 5060 Ti, bf16, in process, median of 30 after warm-up[Vendor claim — not independently verified]
- vLLM server: ~60 requests/s with 8 in flight on one RTX 5060 Ti; 243-intent catalog answered in 0.44 s (vs 4.65 s unpacked)[Vendor claim — not independently verified]
- int8 ONNX on CPU scores 0.891 on razvilka; ~0.9 s for a 384-token request with 3 questions on 6 threads[Vendor claim — not independently verified]
News covering this model
Read this first
- The headline benchmark is the authors’ own. razvilka was published the same day as the model, its scorer lives in the FRIDA-Decisions repo, and every row in the table was run by the team. Treat 0.893 vs Jev 0.897 as self-reported.
- Russian first. Training data is 71% Russian, 29% English, and razvilka is Russian-only. There is no English benchmark on the card.
- Not a
/v1/systemoneserver. The request JSON is Jev-shaped (state,questions,criteria,noul), but serving goes through Python (Judge) or a vLLM pooling plugin. Jev clients need an adapter.
FRIDA-Decisions is an open decision model published on Hugging Face on 2 October 2026 by УЭСМО, SberAI under the ai-forever organization. It is built on ai-forever/FRIDA, a T5 encoder with 823M parameters, and answers structured questions about a text in a single encoder pass, with no generation. Code is on GitHub (MIT, release v0.2.0).
It is the first encoder-only decision model in the catalog built for Russian, and one of the few that runs usefully on CPU: an int8 ONNX build ships with the weights.
Specs
| Attribute | Value |
|---|---|
| Base | ai-forever/FRIDA (T5 encoder, 823M) |
| Decision part | LoRA r16 on q/k/v/o of all 24 layers + scalar head (4.7M trainable), merged into the released weights; head.safetensors separate |
| Question types | choice, score, noul, plus ranking (order candidates) |
| Packing | All options of all questions share one sequence; state cache makes follow-up questions cost only their own tokens |
| Limits | Texts up to 512 tokens recommended; up to 16 options per row |
| Serving | pip install "frida-decisions[torch] @ git+…@v0.2.0" → Judge; [onnx] → OnnxJudge on CPU; vllm serve … --io-processor-plugin frida_decisions → POST /pooling |
| Training | 1.42M examples over 151 question types, 71% Russian / 29% English; ~a quarter LLM-generated or model-labelled (team’s estimate); one epoch, ~50 GPU-hours on an RTX 5060 Ti class card |
| License | MIT (base FRIDA is MIT too) |
| Launch |
Evidence (UnverifiedClaim)
From the model card:
| Model | razvilka accuracy |
|---|---|
| TypeSafe Jev (API) | 0.897 |
| FRIDA-Decisions | 0.893 |
| FRIDA-Decisions, int8 ONNX on CPU | 0.891 |
| smolnikov/migom-2b | 0.853 |
| Mapika/decider-2b | 0.833 |
| fastino/GLiNER2.5-multi-Decide | 0.576 |
| convaiinnovations/laya (multilingual) | 0.559 |
| open-jev (DeBERTa-v3-large) | 0.490 (input window too short for 140 of 735 items) |
| chance | 0.257 |
razvilka has 735 Russian items over 15 tasks (routing, intents, topic, sentiment, moderation, relevance ranking), all four question types, gold labels taken from published datasets. The items and scorer are public, so the table can be rerun, but nobody outside the team has done so yet.
Latency, single stream, in process: 28.2 ms (one question) and 34.0 ms (three questions) on an RTX 5060 Ti; 0.88 s on 6 CPU threads with the int8 build. Through vLLM: 25–45 ms per request and about 60 requests/s with 8 in flight.
Fit / anti-fit
Fit: Russian-language routing, intent detection, moderation and RAG relevance checks; CPU-only or consumer-GPU deployments; large label catalogs (packing makes 243 intents cost 0.44 s instead of 4.65 s).
Anti-fit: English-first workloads until there is an English benchmark; long documents (512 recommended); anything that needs a hosted API, a /v1/systemone endpoint or calibrated probabilities you can take at face value.
Limits
- No calibration study: the card shows distributions but no ECE or temperature.
- Not on the Decision Index or JevBench boards.
- The GitHub repo’s license field reads NOASSERTION; the LICENSE file is MIT (FRIDA base credited).
Why it is in the catalog
FRIDA-Decisions trains its own decision weights (LoRA plus a scalar readout head on an encoder), publishes them with a documented typed API that returns a probability per option, and covers all three decision types. It is listed as model because the release ships merged full weights, not a separate adapter.