CommunityJev Models & Reimplementations
JevK5 — huggingface.co
alibiserikbay's Apache-2.0 Jev alternative on Qwen3.5-4B with a merged rank-16 LoRA: one forward pass returns a probability per option, 0.784 on JevBench's public hard tier at about 13 ms per decision.
- Published by
- Community
- Author
- alibiserikbay
- Last verified
- Oct 2026
This is one of several Jev listings for the same project. The canonical entry is Jevk5 — allebee.
- Languages
- Python
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- Qwen3.5-4B with a rank-16 LoRA on attention projections merged into the weights, read through the SemIf protocol at temperature 1.22.
- Trained on one epoch over 47,460 rows: 17,408 teacher questions and 30,052 human-labelled items from 26 public datasets' train splits.
- JevBench v1.2 public: easy 1.000, standard 0.944 and hard 0.784, with 231/231 valid answers and zero failures.
- Index proxy rose from 0.620 in v0.2 to 0.731; held-out teacher-question accuracy 0.834 with calibrated ECE 0.035.
- Runtime 0.3.0 gives about 13 ms per decision on an H100 via one CUDA graph per padded input length.
Watch out
The index proxy is the author's own estimate, not an index score, and each of its 16 sources has only 40 rows, so any single source is noisy at about ±0.15. It is not Jev's architecture and is not affiliated with TypeSafe AI, and generic generate() calls do not perform the decision readout — the JevK5 runtime is required.
Questions about this entry
Short answers drawn from the linked project or page.
A probability for every option in one forward pass with zero generated tokens, for yes/no (noul), choice and score questions. The readout is a softmax over the answer letters' next-token logits, divided by one calibration temperature.
Generic text-generation examples call generate() and do not perform JevK5's decision readout. Use the JevK5 runtime at github.com/allebee/jevk5 version 0.3.0 or later, which also serves a TypeSafe-style /v1/systemone endpoint.
Five times the teacher data: 17,408 questions from two teachers (Qwen3.6-27B and GPT-6 Luna) instead of 3,272, plus 30,052 human-labelled items from 26 datasets' train splits. It uses no test or validation split and no MMLU, MMLU-Pro or BBH.
231/231 valid answers with zero failures on the v1.2 public items: easy 1.000 (48 items), standard 0.944 (72) and hard 0.784 (111). The hard tier fixed 10 items and broke 5 against v0.2 (McNemar p = 0.30).
Runtime 0.3.0 reads groups of up to 16 options with the knockout readout and combines them through a second temperature, 0.93 in jevk5_config.json. Below 16 options, runtimes 0.2.2 and 0.3.0 give probabilities bit-identical on all 231 public JevBench items.
About 13 ms per decision on an H100 in-process at batch 1 with CUDA graphs: p50 13.2 ms and p95 14.6 ms on easy and standard items, and p50 30 ms with p95 162 ms on hard items carrying 1-4k-token documents.
Apache-2.0. It is an independent open-source alternative to TypeSafe's Jev for typed decisions, not Jev's model or architecture, and the card says it is not affiliated with TypeSafe AI.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
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 NewsHacker News: A one-token logprob wrapper that also classifies webcam frames.
allanrbo151 pointsA single function Jev-like wrapper for LLMs, including vision modelsRead the thread on Hacker NewsRelated terms
Glossary definitions related to this entry.
More Jev resources like this
The 9B JevK5: the same v0.3 data on Qwen3.5-9B, retrained for 1.5 epochs by a pre-registered sweep, reaching index proxy 0.794 and a 0.859 hand-written hard set, Apache-2.0.
Jev Models & Reimplementations#community#open-models#local
Community
AutoTrust's combined System 1 and System 2 bundle on Gemma-4-26B-A4B-it: Decision Index 0.2.1 of 58.05, a 24-slot head with noul, choice and score, and the base chat model left bit-identical.
Jev Models & Reimplementations#community#open-models#local
Community
GGUF builds of the JevK5 family for llama.cpp: Q8_0 is effectively lossless (229 of 231 answers match the 4B v0.3 bf16), and every published file clears the card's 95% agreement gate.
Jev Models & Reimplementations#community#open-models#local
Community
chaoliangUNSW's v1 Jev-Style decision model in GGUF from 1.3 GB: calibrated choice, bool and score answers on Qwen3.5-2B-Base, 82.3% on five held-out task families with ECE 0.017.
Jev Models & Reimplementations#community#open-models#local
Community
An Apple Silicon MLX bf16 build of the v1 Jev-Style decision model: 77 ms decisions on an M1 Max with calibrated choice, bool and score probabilities, Apache-2.0 weights.
Jev Models & Reimplementations#community#open-models#local
Community
Mithril's open, Jev-shaped typed-decision model on DeBERTa-v3-large: choice up to 255 options, score and noul from one forward pass, 0.854 in-domain with ECE 0.022, trained on public gold labels only.
Jev Models & Reimplementations#community#open-models#local
Community
From the Jev guides
Original write-ups on what this entry is for.
Training recipes, looped architectures, and language ports: what the open side of the typed-decision class shipped this month, and how to pick by runtime and calibration instead of the leaderboard.
Jev Models & Reimplementations#open-models#training#local
Read article
Taking a decision checkpoint out of Python: what the GGUF and browser WebGPU exports preserve, what they change, and the parity evidence the model cards publish.
Jev Tools & CLIs#open-models#local#onnx
Read article
Interface, size, license, claims — in that order. A selection guide across the system-one catalog, from a 706K form scorer to 27B Jev students.
Jev Practices & Patterns#open-models#evaluation#benchmarks
Read article
From the Jev community
Posts from builders shipping with Jev right now.