Article
CommunityJev Models & Reimplementations
TrialJev
Open-weight eligibility scorer that LoRA-adapts Qwen models with a pointer head to score natural-language answer options in one forward pass, trained on TREC Clinical Trials 2021 and evaluated on 6,975 pairs from 50 unseen synthetic patients.
- Format
- Article
- Reading time
- 2 min
- Published by
- Community
- Added
- Nov 2026
Highlights
- Scores natural-language answer options in a single forward pass with no autoregressive decoding.
- Reaches 80.97% (Qwen3.5-27B) and 81.84% (Qwen3.8-27B) accuracy on 6,975 eligible-or-excluded pairs.
- Verdict generation takes 2.45-2.70 times as long as one-pass inference at 9B-27B in matched batch-1 measurements.
- Runs locally on a single GPU at roughly 178 ms per pair at 27B, without sending patient text to an external API.
- Adapters, pointer heads, and model cards are released on Hugging Face.
Questions about this entry
Short answers drawn from the linked project or page.
The paper reports that adapters, pointer heads, and model cards are released on Hugging Face. TrialJev runs locally on a single GPU and requires approximately 178 ms per pair at 27B, without transmitting patient text to an external model API.
On 6,975 eligible-or-excluded pairs from 50 unseen TREC 2022 synthetic patients, TrialJev achieves 80.97% accuracy with Qwen3.5-27B and 81.84% with Qwen3.8-27B.
Under a common development-fitted decision correction, the two 27B TrialJev models reach 80.99% and 81.88% accuracy, compared with 81.28% for the proprietary hosted Jev API. The paired confidence intervals include zero.
TrialJev is trained on TREC Clinical Trials 2021 and evaluated on 6,975 eligible-or-excluded pairs drawn from 50 unseen TREC 2022 synthetic patients.
No. TrialJev runs locally on a single GPU, so patient text is not transmitted to an external model API. The paper frames this as useful under restrictions on sharing patient records.
An A/B-logit ablation shows that avoiding decoding, rather than the pointer head alone, accounts for the measured efficiency advantage over LLMs fine-tuned to generate verdicts.
Evaluation is limited to synthetic patients and implementation-specific timing, and clinical use has not been validated. TrialJev is intended for first-pass prioritization of candidate trials, complementing rather than replacing explanatory models and clinical review.
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
LoRA adapter set for the Jevvideo long-video frame-selection agent, trained on 8xH200 from OmniJev/OneJev-4B. Ships controller, grounding, and verifier adapters plus per-role metrics and five-benchmark results.
Jev Models & Reimplementations#open-models#video#benchmark
Community
arXiv paper presenting EDU-Jev, a family of single-pass System One decision models at 0.8B, 2B and 4B parameters that return a probability distribution over predefined education answer options without generating any text.
Jev Models & ReimplementationsArticle#arxiv#open-models#classification
Community
arXiv paper introducing Drex, a family of four open decision models at 8B, 9B, 35B with 3B active, and 27B parameters. Reports 67.52 on Decision Index 0.4.1 versus 60.70 for Jev 1.14.0.
Jev Models & ReimplementationsArticle#arxiv#open-models#benchmark
Community
Zuse is a family of three decision models (4B, 9B, 27B) that train a low-rank adapter and pointer head on frozen open-weight bases for 27 European languages. The 27B model reaches 87.5% accuracy with pooled expected calibration error of 0.016 on an 11-source European suite.
Jev Models & ReimplementationsArticle#arxiv#open-models#calibration
Community
Fine-tuning method that turns fix commits into comparison groups so an open-weight model scores code functions for vulnerability. An ensemble of five Qwen3.5-4B models reaches 0.814 AUROC on a 7,270-function benchmark, against 0.743 for Jev 1.13.
Jev Models & ReimplementationsArticle#arxiv#open-models#security
Community
arXiv paper introducing Flide, an exact inference engine for decision models such as Jev. It decouples the logical tree of a decision request from physical execution and reports answers up to 9.5 times faster than vLLM and SGLang.
Jev Models & ReimplementationsArticle#arxiv#benchmark#evaluation
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.