Jev directory
CommunityJev Models & Reimplementations
Jev Boost re6-hover-timeqa Embeddings (Qwen3.6-35B-A3B-FP8)
Hugging Face dataset of Jev Decision Index-shaped state/action embeddings built from TimeQA. Holds 13,278 train and 1,460 validation pairs embedded with Qwen3.6-35B-A3B-FP8 at 2048 dimensions for post-training CLM heads.
- Reading time
- 2 min
- Published by
- Community
- Added
- Oct 2026
Highlights
- 13,278 train and 1,460 validation pairs, each paired with one hard negative (k=1).
- Embeddings are float16 [n, 2048] from vLLM last-token pooling, L2-normalized.
- States render the target benchmark's request and instructions; options are never listed in the state.
- Items sharing a paragraph with a Jev suite item are dropped to avoid overlap.
- One-token actions and negatives are zeroed and resolved from a separate single-token table.
Questions about this entry
Short answers drawn from the linked project or page.
It supplies state/action embeddings for post-training CLM heads through clm_train.data.jev_boost. Records are converted into decisions shaped like the Jev Decision Index's Retrieval & Classification requests, using HoVer claim verification as the target benchmark.
Qwen/Qwen3.6-35B-A3B-FP8, run with vLLM pooling on the last token and L2-normalized to 2048 dimensions. The metadata records max_model_len 16,384 and state_head_tokens 2,048.
13,278 train pairs and 1,460 validation pairs, each with one hard negative. The build caps at 20,000 rows using a hash sample, and the Hugging Face viewer shows two rows.
Hard negatives are the other options for a record, at most 8 per row, with near-copies of the gold option dropped. This build uses k=1; neg_embeddings.pt is zero-padded and neg_counts.pt gives the real count per row.
One-token sequences are never embedded, because vLLM's pooling of a one-token prompt is noise for this model. They are zero here, and single_tokens.npz points them at jackyk02/qwen3.6-35b-a3b-fp8-single-token-embeddings, which clm_train.data.stores.EmbeddingDir applies on load.
The source's license: BSD-3-Clause-Clear for TimeQA and CC BY-SA for the Wikipedia text. The embedding vectors are a derivative of that text, so the same terms carry over.
Use clm-train posttrain with the source hf:boost-re6-hover-timeqa-qwen3.6. Each split directory holds state_embeddings.pt, action_embeddings.pt, neg_embeddings.pt, neg_counts.pt, single_tokens.npz and metadata.json; total file size is 188 MB.
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 that converts AVeriTeC claim-verification records into Jev Decision Index retrieval and classification requests, embedded with Qwen3.6-35B-A3B-FP8 into 2048-d vectors for post-training CLM heads.
Jev Models & Reimplementations#open-models#classification#benchmark
Community
Hugging Face dataset of Jev Decision Index-shaped state/action embeddings built from Re-DocRED. Holds 13,312 train and 1,517 validation pairs embedded with Qwen3.6-35B-A3B-FP8 at 2048 dimensions for post-training CLM heads.
Jev Models & Reimplementations#open-models#classification#benchmark
Community
Hugging Face dataset of Jev-shaped retrieval and classification decisions built from MM-BRIGHT, holding Qwen3.6-35B-A3B-FP8 state, action, and hard-negative embeddings for post-training CLM heads.
Jev Models & Reimplementations#open-models#benchmark#evaluation
Community
Hugging Face dataset that renders HoVer claim-verification items as Jev Decision Index retrieval and classification requests, with Qwen3.6-35B-A3B-FP8 state, action, and hard-negative embeddings for post-training CLM heads.
Jev Models & Reimplementations#classification#benchmark#open-models
Community
Hugging Face dataset of Qwen3.6-35B-A3B-FP8 embeddings over PRESTO's English tool-use requests, shaped as Jev Decision Index When2Call multiple-choice items for post-training CLM heads.
Jev Models & Reimplementations#open-models#benchmark#classification
Community
A family of 0.8B-9B ophthalmic decision models that score candidate answers without autoregressive decoding, trained on RetiJev-Data: 78,158 clinical-decision and diagnosis questions and 50,074 guideline-grading questions. RetiJev-9B answers a case in 158 ms on one GPU.
Jev Models & ReimplementationsArticle#arxiv#open-models#benchmark
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.