Jev directory
CommunityJev Models & Reimplementations
Jev Boost: To6-Presto-When2Call Qwen3.6-35B-A3B-FP8 Embeddings
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.
- Reading time
- 2 min
- Published by
- Community
- Added
- Oct 2026
Highlights
- 18,038 train and 1,962 validation pairs, each paired with 3 hard negatives.
- Vectors are 2048-d float16 from vLLM last-token pooling, L2-normalized.
- States carry the benchmark's request plus its own instructions and never list the options.
- 33 intents as tools in 7 sibling families, labeled call, request for info, or cannot answer.
- One-token sequences are zeroed and resolved from a separate single-token embedding table.
Questions about this entry
Short answers drawn from the linked project or page.
It is consumed through clm_train via clm_train.data.jev_boost, documented in docs/jev-boost.md; the published command is clm-train posttrain --source hf:boost-to6-presto-when2call-qwen3.6.
Qwen/Qwen3.6-35B-A3B-FP8, embedded with vLLM pooling using the last token and L2 normalization, producing 2048-dimensional float16 vectors.
fathyshalab/google-presto at revision 5747a2076518fb1a66d39f2af1fb5011e4be4327, train split, with English requests kept including disfluencies and in-sentence corrections, covering 33 intents as tools in 7 sibling families.
The license is per-source: the PRESTO source is CC BY 4.0 (Google), and the released vectors are a derivative of that source text.
Negatives are the other options of each item, at most 8 per row, with near-copies of the gold option dropped; the released splits carry k=3 negatives stored as a zero-padded [n, k, 2048] tensor with per-row counts.
vLLM pooling of a one-token prompt is noise for this model, so one-token sequences are never embedded; single_tokens.npz points them at jackyk02/qwen3.6-35b-a3b-fp8-single-token-embeddings, applied on load by clm_train.data.stores.EmbeddingDir.
The vectors are tied to Qwen3.6-35B-A3B-FP8 and will not match other embedding models; it covers English PRESTO requests only, items whose state shares a paragraph with a Jev suite item are dropped, and rows are capped at 20,000 by hash sample.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
Hacker News: ASTGL replay chose local Laya routing over hosted Jev on one Mac Studio setup.
Jmeg8r15 pointsDoes an open-weight decision model beat a hosted one? Jev vs. LayaRead the thread on Hacker NewsReddit: A walkthrough of routing between models with Jev.
X: Standard One SH (StandardThinking)
Facebook: Intake Router keeps diagnosis out of the answer space at the front desk.
FacebookClinicnotes
More 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 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.
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
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.