Jev directory
CommunityJev Models & Reimplementations
Jev Knowledge Qwen3.6-35B-A3B-FP8 Embeddings (256k)
Hugging Face dataset of Qwen3.6-35B-A3B-FP8 state/action embeddings for post-training contrastive heads on the Jev Decision Index's Knowledge & Reasoning train splits: CommonsenseQA, MMLU auxiliary_train and ChessBench.
- Reading time
- 2 min
- Published by
- Community
- Added
- Oct 2026
Highlights
- 355,093 train and 39,554 validation (state, action) pairs at 2048 dimensions
- Built from CommonsenseQA, MMLU auxiliary_train and ChessBench train splits
- Up to 8 hard negatives per row, zero-padded, with valid counts in neg_counts.pt
- Two state views per decision item: the state alone and the state with its options listed
- One-token actions, negatives and states take vectors from a separate single-token table
Questions about this entry
Short answers drawn from the linked project or page.
The card gives python scripts/download.py data jev-knowledge-qwen3.6 --rows 20000 --blocks 8, and a clm-train posttrain --source hf:jev-knowledge-qwen3.6 command. The store is also readable through the Datasets library and the Hub's dataset viewer.
Qwen/Qwen3.6-35B-A3B-FP8, run through a vLLM pooling runner with last-token pooling and L2 normalization, producing 2048-dimensional vectors with max_model_len 8,192.
The dataset is licensed per-source. The vectors are a derivative of the source text and carry that text's terms, which are listed in the dataset documentation.
The split is by task, not by row: int(task_id[:8], 16) % 10 == 0 sends a task to validation. Row i of every tensor and samples[i] in metadata.json describe the same pair.
vLLM's pooling of a one-token prompt returns noise for this model, so every one-token action, negative or state takes its vector from the jackyk02/qwen3.6-35b-a3b-fp8-single-token-embeddings table. single_tokens.npz lists those rows and token ids.
MMLU-Pro, GPQA Diamond and MuSR are skipped because they have no train split or are evaluation-only. Validation and test items of the scored benchmarks are never read, and the ChessBench test bag is never downloaded.
Each split holds float16 .pt tensors for state, action, negative and options-view embeddings, int64 neg_counts.pt, bool has_options.pt, single_tokens.npz and metadata.json. Tensors load with torch.load(..., mmap=True).
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
Hugging Face dataset of Qwen3.6-35B-A3B-FP8 state/action embeddings for contrastive post-training on the Jev Decision Index's entailment and verification train splits: ANLI, ContractNLI, NLI4CT, HoVer and RAGTruth.
Jev Models & Reimplementations#open-models#calibration#research
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
SecJev is a family of security-specialized System One decision models from 0.8B to 9B parameters, built on Kev's single-pass candidate scorer. It spans 14 tasks and eight sources, and SecJev-0.8B beats general Kev-9B by 20.51 points in task-macro accuracy.
Jev Models & ReimplementationsArticle#community#research#arxiv
Community
Shows that open instruction-tuned models already read decisions from answer-letter logits without training, and that LoRA fine-tuning with a decision head lifts them past Jev: Gemma 4 31B goes from 79.6 to 84.4 on the new 2,352-question OpenDecisions set and scores 81.5 on JevBench v1.7.
Jev Models & ReimplementationsArticle#community#research#arxiv
Community
A System One model for Chinese-language decisions trained on a 10-million-example corpus and fine-tuned for medicine, legal, and finance. Pre-training exceeds closed-source Jev by 1.24% on general tasks with a 20.3x speedup; domain fine-tuning beats Jev in medicine by 4.0%.
Jev Models & ReimplementationsArticle#community#research#arxiv
Community
1.4kGitHub stars
Reads a typed decision from one prefill of a pretrained instruction-tuned model by restricting the next-token distribution to option tokens, then corrects label and position bias with no gradient steps. Averaging over cyclic rotations lowers order-flip rates and raises accuracy on 11 of 11 models.
Jev Models & ReimplementationsArticle#community#research#arxiv
Community
nokia-applied-research/AnyJevPython
nokia-applied-research/AnyJevPythonFrom 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.