Jev directory
CommunityJev Models & Reimplementations
Jev Retrieval Qwen3.6 35B A3b Fp8 Embeddings 256K
Hugging Face dataset of (state, action) embedding pairs from Qwen3.6-35B-A3B-FP8, rendered from the ESCI, ToolRet and API-Bank train splits for post-training contrastive retrieval heads. Ships float16 tensors, hard negatives and per-row metadata.
- Reading time
- 2 min
- Published by
- Community
- Added
- Oct 2026
Highlights
- 2048-d L2-normalized last-token embeddings from Qwen/Qwen3.6-35B-A3B-FP8 via a vLLM pooling runner
- Train split holds 644,728 pairs across 273,885 tasks; validation holds 72,709 pairs
- Each row carries k=7 hard negatives, zero-padded, with neg_counts.pt giving valid counts
- Decision items store two state views: the state alone and the state with its options listed
- One-token actions, negatives and states take vectors from a separate single-token embedding table
Questions about this entry
Short answers drawn from the linked project or page.
The card documents clm-train posttrain --source hf:jev-retrieval-qwen3.6, or python scripts/download.py data jev-retrieval-qwen3.6 --rows 20000 --blocks 8.
Qwen/Qwen3.6-35B-A3B-FP8, run through a vLLM pooling runner with last-token pooling and L2 normalization, producing 2048-dimensional vectors.
The train splits of Amazon ESCI, ToolRet and API-Bank, rendered with the same fields, order, instruction and option set that the index's builder uses for a scored item.
neg_embeddings.pt is float16 [n, k, 2048] with k=7, zero-padded; neg_counts.pt is int64 [n] giving the number of valid negatives per row.
The split is by task: int(task_id[:8], 16) % 10 == 0 routes a task to validation, and the remaining tasks go to train.
vLLM's pooling of a one-token prompt returns noise for this model, so one-token actions, negatives and states take vectors from jackyk02/qwen3.6-35b-a3b-fp8-single-token-embeddings, listed in single_tokens.npz.
The card lists the license as per-source. The vectors are a derivative of the source text and carry its terms, so downstream use inherits the source datasets' conditions.
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
PostHog's reasoning Jev-style classifier with a diffusion drafter, reporting higher accuracy than Kev-9B and hosted Jev on held-out and JevBench tiers.
Jev Models & Reimplementations#community#evaluation#open-models
Community
PostHog/jeevesPython
PostHog/jeevesPythonA preregistered measurement of a Jev-style interface on frozen Qwen3-4B: read the option letter's logits instead of generating grammar-constrained JSON.
Jev Models & Reimplementations#community#python#local
Community
r-ms/mini-jevPython
r-ms/mini-jevPythonOpen Jev-style decision model built on LiquidAI's LFM2.5-350M backbone, a clone of kev. Serves the System One API contract so the TypeSafe SDK works against a local server, returning probabilities for typed questions from one forward pass.
Jev Models & Reimplementations#open-models#python#local
Communityfranckverrot/lev
Hugging Face dataset of (state, action) embedding pairs from Qwen/Qwen3.6-35B-A3B-FP8 for post-training contrastive heads on the Jev Decision Index's Arts & Human Taste benchmarks: Humicroedit, cfcolor and the Habermas Machine.
Jev Models & Reimplementations#open-models#research#benchmark
Community
Hugging Face dataset of (state, action) embedding pairs from Qwen/Qwen3.6-35B-A3B-FP8 for post-training contrastive heads on the Jev Decision Index's Language Understanding benchmarks: iSarcasmEval, VAST, ACOS and New Yorker caption matching.
Jev Models & Reimplementations#open-models#research#benchmark
Community
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
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.