Jev directory
CommunityJev Models & Reimplementations
Turnike-JEV-9B
A 9B bilingual Turkish and English Jev-class decision model from Berkay Hasret: a 173 MB LoRA adapter on Qwen/Qwen3.5-9B that answers up to 16 typed questions about one text in a single forward pass and reports 0.783 macro accuracy on full JevBench.
- Reading time
- 3 min
- Published by
- Community
- Author
- berkay123001
- Last verified
- Oct 2026
Used inJev for Software & DevTools
- Languages
- Python
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- LoRA adapter (r=16, 43.3M trainable parameters, about 173 MB) on Qwen/Qwen3.5-9B; up to 16 questions about one text in one pass.
- Reports 0.783 macro accuracy and 0.076 macro ECE on the full JevBench test (22 configs, 22,773 rows), with 0.842 on six held-out sources.
- On 1,379 human-labelled Turkish decisions it reports 0.763, ahead of Microsoft-Decision-1 at 0.714 (paired McNemar p = 2e-6).
- FP8 fits a 16 GB GPU at about 9 GB steady state and 12 GB peak loading; about 139 ms per decision on an RTX 5060 Ti, single run.
- Not distilled from a commercial decision API; training data not released; card and code written with AI assistance.
Watch out
Card reviewed 2026-10-17; the 0.783 JevBench macro accuracy, the 0.842 held-out result, the Turkish 0.763 and the latency and memory figures are self-reported in a single run and were not submitted to the public leaderboard. JevBench here means the Praveenrajus/jev-bench dataset scored with jevify-run, not the Benchmark Heaven board of the same name.
Questions about this entry
Short answers drawn from the linked project or page.
A bilingual Turkish and English Jev-class decision model from Berkay Hasret: a 173 MB LoRA adapter (r=16, 43.3M trainable parameters) on Qwen/Qwen3.5-9B. Text and closed questions go in; option probabilities come out, fully offline.
Text and questions go into one chat prompt whose assistant turn is a JSON skeleton with one placeholder per question; each option maps to a single-token symbol, and a softmax over those symbols' output rows at the placeholder position yields the probabilities.
Macro accuracy 0.783 and macro ECE 0.076 over 22 configs and 22,773 rows, with 0.842 on six held-out sources and 0.761 on the other 16. The card notes this is a single run, not submitted to the leaderboard whose top entry it reports at 0.770.
On 1,379 human-labelled Turkish decisions it reports 0.763 versus 0.727 for Jev 1.13.0, 0.714 for Microsoft-Decision-1 (p = 2e-6) and 0.708 for Kev-4B. On machine-translated TR-50 it reports 0.721, behind Microsoft-Decision-1 at 0.745.
A LoRA adapter of about 173 MB on Qwen/Qwen3.5-9B, or the merged bf16 or stored FP8 repositories. The FP8 copy fits a 16 GB GPU at about 9 GB steady state and 12 GB peak, with roughly 139 ms per decision and 206 ms for eight questions.
Both are post-trained Qwen3.5-9B models, but Microsoft-Decision-1 is closed-weight and API-only while Turnike-JEV-9B runs locally. The card reports Turnike ahead on Turkish human labels (0.763 vs 0.714) and level on JevBench EN-300 (0.780 vs 0.775).
Apache-2.0, with the base Qwen/Qwen3.5-9B under Apache-2.0. Training used public Turkish and English datasets, LLM-labelled public texts and synthetic data that are not released; the card says the model was not distilled from any commercial decision API and that no web data is in this release.
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
A bilingual Turkish and English 4B Jev-class decision model from Berkay Hasret: a roughly 130 MB LoRA adapter on Qwen/Qwen3.5-4B that answers up to 16 typed questions per text in one pass and reports 0.774 JevBench macro accuracy.
Jev Models & Reimplementations#community#open-models#local
Community
A 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
A 24 GB-GPU alternative from the train-jev project: Qwen3.5-4B plus a rank-16 LoRA and 24-slot decision head scoring 0.647 PK accuracy, 40.4 Decision Index and 43 ms p50 latency, ahead of Kev-4B at the same size.
Jev Models & Reimplementations#community#open-models#local
Community
A Laya fine-tune for coding decisions: roughly 9k rows from SWE-Gym teach keep/drop file relevance, priority scoring and injection detection, reporting 0.833 accuracy on 1,150 held-out rows at 126 ms p50.
Jev Models & Reimplementations#community#open-models#local
Community
A Laya fine-tune from Convai Innovations on the 1,200-case typed-decisions benchmark: it reports 0.777 accuracy on the official 400-case test set, above TypeSafe Jev 1.13.0 at 0.727 and the teacher ceiling at 0.735.
Jev Models & Reimplementations#community#open-models#local
Community
Mapika's NVFP4 fine-tune of Gemma-4-31B-it answers typed Choice, Noul and Score questions with calibrated probabilities and no generated text, self-reporting a 64.27 chance-corrected public Decision Index and 39 ms median latency on a B300.
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.