Jev directory
CommunityJev Models & Reimplementations
Jev Boost to6 AirDialogue when2call Qwen3.6-35B-A3B-FP8 Embeddings
A Hugging Face dataset of Qwen3.6-35B-A3B-FP8 state, action, and hard-negative embeddings built from AirDialogue's book, change, and cancel requests, shaped as When2Call multiple-choice decisions for post-training CLM heads.
- Reading time
- 2 min
- Published by
- Community
- Added
- Oct 2026
Highlights
- 10,818 train and 1,182 validation pairs, each stored with 3 hard negatives
- Embeddings are 2048-d float16, L2-normalized, last-token pooled with vLLM
- Covers travellers' book, change, and cancel requests and the agent's next move
- Rows are capped at 12,000 by hash sample; items sharing a paragraph with a Jev suite item are dropped
- Licensed CC BY-NC 4.0 from the AirDialogue source; the vectors are a derivative of its text
Questions about this entry
Short answers drawn from the linked project or page.
The card documents loading through clm_train, which reads it with clm_train.data.jev_boost and the EmbeddingDir store; the documented command is clm-train posttrain --source hf:boost-to6-airdialogue-when2call-qwen3.6. The .pt and .npz files can also be read directly with the Hugging Face Datasets library, pandas, or Polars.
Vectors come from Qwen/Qwen3.6-35B-A3B-FP8 using vLLM pooling with the last token, L2-normalized to 2048 dimensions. Metadata records hidden_size 2,048, max_model_len 16,384, state_head_tokens 2,048, and the clm_train.render.Renderer renderer.
The train split of google/air_dialogue at revision dbdbe7bcef8d344bc3c68a05600f3d95917d6898: travellers' opening book, change, and cancel requests plus the agent's next move when it agrees with the request. Missing dates, airports, or a name become request-for-info; otherwise the search or reservation call.
Train holds 10,818 pairs and validation 1,182, each with 3 hard negatives in neg_embeddings.pt. At most 8 hard negatives per row are considered, near-copies of the gold option are dropped, and neither split contains one-token actions or negatives.
The card states that vLLM's pooling of a one-token prompt is noise for this model, so one-token sequences are never embedded and are zeroed here. single_tokens.npz holds their token ids and points them at jackyk02/qwen3.6-35b-a3b-fp8-single-token-embeddings, applied on load by EmbeddingDir.
The dataset carries the source's license, CC BY-NC 4.0 for the AirDialogue data, and the card notes no license is stated on the source card itself. The vectors are described as a derivative of that source text, so the same non-commercial terms are expected to apply.
Each split provides state_embeddings.pt and action_embeddings.pt as float16 [n, 2048], neg_embeddings.pt as float16 [n, k, 2048] zero-padded, neg_counts.pt as int64 [n], single_tokens.npz, and metadata.json with one dict per row. Total file size is 252 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
A Hugging Face dataset of Qwen3.6-35B-A3B-FP8 state, action, and hard-negative embeddings built from xSID's English alarm, reminder, and weather requests, shaped as When2Call multiple-choice decisions for post-training CLM heads.
Jev Models & Reimplementations#huggingface#open-models#benchmark
Community
Hugging Face dataset of 6,919 API-Bank training decisions rendered in the Jev Decision Index shape and embedded with Qwen/Qwen3.6-35B-A3B-FP8 at 2048 dimensions, for post-training CLM heads.
Jev Models & Reimplementations#huggingface#embeddings#open-models
Community
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
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.