Jev directory
CommunityJev Models & Reimplementations
CLM-35B-A3B Jev Boost Head
A contrastive state/action projection head over Qwen3.6-35B-A3B embeddings, post-trained on 67 text stores plus 382 Jev boost stores. Scores decisions as cosine similarity between state and action heads.
- Reading time
- 2 min
- Published by
- Community
- Added
- Oct 2026
Highlights
- 3-layer GELU MLP head, width 1536, 512-d output over 2048-d embeddings
- Trained on 67 text stores plus 382 Jev boost stores with 40% HQ-DQA replay
- Best checkpoint at step 33,000 of 40,000 by mean store validation loss
- Head text index scores 39.9 overall versus 21.6 for the 67-store baseline
- Built via clm_train with heads.build(2048, "best_head.pt")
Questions about this entry
Short answers drawn from the linked project or page.
It is a contrastive state/action projection head trained over Qwen/Qwen3.6-35B-A3B-FP8 embeddings. It produces a 512-d output from 2048-d L2-normalized embeddings and scores decisions as the cosine similarity between the state and action heads.
Use clm_train.models.heads.build(2048, "best_head.pt") to get a pair object exposing pair.state_head and pair.action_head. Score decisions as cos(state, action).
States and actions must be rendered and embedded exactly as in training, using clm_train.render.Renderer. One-token sequences take the single-token table.
HQ-DQA pre- and mid-training, then post-training on the recipe's 67 text stores plus 382 Jev boost stores, with plain states and 40% HQ-DQA replay. The mixture is in joint_weights.json.
On the Jev Decision Index text public suite (35 of 37 panel benchmarks), the head scores 39.9 overall versus 21.6 for the 67-store baseline, with knowledge 10.9, language 52.1, retrieval 54.7, tools 60.5, and arts 16.5.
Training data licences vary per source, and several sources are non-commercial. The model card points to the boost recipe for details.
No. The model card states it is not deployed by any Inference Provider, so it must be run locally or self-hosted.
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
Satirical Python script that answers a query with DO or DO_NOT plus a random confidence percentage, presented as a natural alternative to Jev. MIT-licensed, Python 3.8+, one dependency, no API keys or model weights.
Jev Models & Reimplementations#github#python#open-models
CommunityAhmad170412/Koin
Russian-first open typed-decision model on Hugging Face. Given a context and a set of candidate options it returns a calibrated probability and picks one, covering yes/no, choice, and ordinal scale decisions. Built on mmBERT-base with 307M parameters under an MIT license.
Jev Models & Reimplementations#open-models#calibration#classification
Community
A 561M-parameter XLM-RoBERTa-large decision model given one extra training epoch on Jev's answers, with about 23% of examples black-box distilled from Jev's public API. Supports noul, choice, and score questions in 18 languages and loads through the bundled naluri.py.
Jev Models & Reimplementations#open-models#calibration#benchmark
Community
Hugging Face dataset repository holding four supervised fine-tuning configs for JEV decision models: AHAT and PARTNR next-action choices, plus AHAT and PARTNR progress choices over a fixed 21-option grid. Gated access; trainer records and audits only, no checkpoints.
Jev Models & Reimplementations#open-models#calibration#python
Community
A Hugging Face dataset of Qwen3.6-35B-A3B-FP8 embeddings for ToolRet training records, converted into Jev Decision Index ToolRet-style query-to-document decisions for post-training CLM heads. Ships float16 state, action, and hard-negative vectors plus per-row metadata.
Jev Models & Reimplementations#open-models#huggingface#jev-decision-index
Community
An independent, from-scratch reconstruction of a Jev-style decision model that never generates text: a state plus typed questions returns a probability for every option. Trained on Qwen3-0.6B and 1.7B backbones with LoRA and served through a system_one-shaped API.
Jev Models & Reimplementations#open-models#python#research
CommunityES7/JEV-from-Scratch
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.