Jev directory
CommunityJev Models & Reimplementations
FRIDA-Decisions
FRIDA-Decisions answers Russian text with structured decisions in a single encoder pass — choice, ordinal score, yes/no or ranking — and scores 0.893 on the 735-item razvilka benchmark, effectively level with TypeSafe Jev's 0.897 (McNemar p = 0.84).
- Reading time
- 3 min
- Published by
- Community
- Author
- AI Forever
- Last verified
- Oct 2026
- Platforms
- API
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- One encoder pass over Russian text: choice among K options, ordinal score, yes/no and ranking, with every answer coming from the declared options.
- Built on ai-forever/FRIDA (T5 encoder, 823M parameters) by УЭСМО, SberAI; runs on consumer GPUs, with int8 ONNX CPU and MLX builds.
- Scores 0.893 on the 735-item razvilka benchmark; TypeSafe Jev, a commercial API, scores 0.897 on the same items (paired McNemar p = 0.84).
- 28–34 ms per request on an RTX 5060 Ti; a 243-intent catalog is answered in 0.44 s, and vLLM serves about 60 requests/s with 8 in flight.
- New label sets are new JSON, not new training runs: options are written as text inside the request, with no generation, no output tokens and no parsing.
Quickstart
bash
pip install "frida-decisions[torch] @ git+https://github.com/ai-forever/[email protected]"Watch out
Card reviewed 2026-10-17; all benchmark and latency figures are self-reported by the authors. The 0.893 razvilka score is within noise of TypeSafe Jev's 0.897 (paired McNemar p = 0.84), the int8 ONNX build scores 0.891 and matches the GPU model on 726 of 735 items, and no result beyond razvilka is published, so broad generalization is not established.
Questions about this entry
Short answers drawn from the linked project or page.
An encoder decision model from AI Forever (developed by УЭСМО, SberAI) that makes structured decisions over Russian text in one pass: pick one of K options, place text on an ordinal scale, answer yes/no or rank candidates, returning the declared option with its confidence.
The card reports 0.893 on the 735-item razvilka benchmark against 0.897 for TypeSafe Jev on the same items, with a paired McNemar p = 0.84, which the card presents as statistically level, and notes it is the highest score among the open models they ran.
Choice (pick one of K text options), score (an ordinal scale), noul (yes or no) and ranking of candidates. Options are written as text inside the request, so a new label set is a new JSON request rather than a new training run.
28–34 ms per request on an RTX 5060 Ti for a roughly 400-token text with 1–3 questions, 1.8 GiB peak PyTorch allocation over the razvilka run, an int8 ONNX build on CPU scoring 0.891, and an MLX build for Apple Silicon.
pip install the torch or onnx extra from the GitHub v0.4.0 tag, then Judge.from_pretrained('ai-forever/FRIDA-Decisions', device='cuda') and call judge.judge(request). A vLLM server path with the frida_decisions io processor is also documented for many-user serving.
The card models Russian text and its metadata lists ru and en; the quickstart and benchmark focus on Russian, including the razvilka suite. Nothing in the card documents performance for other languages.
MIT licence, built on ai-forever/FRIDA, a T5 encoder with 823M parameters, and developed by УЭСМО, SberAI. A Habr article, GitHub repository with training and serving code, a Spaces demo and quickstart notebooks are linked from the card.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Standard One SH (StandardThinking)
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
Hacker News: Decision Index 0.2.1 scores open reproductions of Jev on a frozen suite.
jackbravo4 pointsDecision Index 0.2.1 – 70 open reproductions of TypeSafe Jev's Decision ModelRead the thread on Hacker NewsHacker News: Pointer to a DiffusionGemma plus vLLM tutorial for a Jev-shaped classifier.
1 pointHow to build a Jev-style classifier with DiffusionGemma and vLLMRead the thread on Hacker NewsMore Jev resources like this
The v2.2 PEFT LoRA adapter for Standard Thinking's Standard One 3B decision model, shipped with a merge recipe for Ministral-3-3B-Instruct; serving requires the merged checkpoint and the shared server code in the StandardOne-8B repository.
Jev Models & Reimplementations#community#open-models#local
Community
The Ollama build of iapp's OpenThai-SystemOne: a 0.8B Thai and English decision model whose v0.3 weights are fine-tuned for Ollama's /v1/systemone letter readout, reporting a 74.3 macro across 13 public subsets for Q8_0.
Jev Models & Reimplementations#community#open-models#local
Community
Standard Thinking's merged BF16 3B checkpoint of Standard One: it scores choice, noul and score questions and returns probabilities through POST /v1/systemone with no generated text. The v2.2 card reports 100.00% JevBench public easy and 79.97% on many-option questions.
Jev Models & Reimplementations#community#open-models#multilingual
Community
A local server that reads next-token probabilities from a fine-tuned Spark model and returns Choice, Score, and Noul answers with zero generated tokens.
Jev Models & Reimplementations#community#python#open-models
Community
Rizzo-AI-Academy/rizzo-flowPython
Rizzo-AI-Academy/rizzo-flowPythonA local /v1/systemone server that reads label probabilities from a GGUF chat model, with prompt repetition and a confidence shrink toward uniform.
Jev Models & Reimplementations#community#python#open-models
Community
Mushroom-Systems/lichenPython
Mushroom-Systems/lichenPythonA LoRA plus 24-slot decision head over Qwen3.8-27B from Ming Xu's train-jev project: it serves TypeSafe's System One API with 0.851 Kev transfer-v4 accuracy and ECE 0.053, and needs an 80 GB GPU.
Jev Models & Reimplementations#community#open-models#local
Community
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.