Jev directory
CommunityJev Models & Reimplementations
Jev Boost to7-w2ctrain-when2call Qwen3.6-35B-A3B-FP8 Embeddings
Embedding dataset from clm-project that renders NVIDIA's When2Call training split into Jev Decision Index-style tool-use decisions and stores Qwen3.6-35B-A3B-FP8 state, action, and hard-negative vectors for post-training CLM heads.
- Reading time
- 2 min
- Published by
- Community
- Added
- Oct 2026
Highlights
- Stores 2048-d L2-normalized embeddings from Qwen/Qwen3.6-35B-A3B-FP8 via vLLM last-token pooling.
- Layout: state_embeddings.pt, action_embeddings.pt, neg_embeddings.pt, neg_counts.pt, single_tokens.npz, metadata.json.
- Layout describes 16,510 train pairs and 1,856 validation pairs, each with k=3 hard negatives.
- One-token actions and negatives are zeroed and resolved from a separate single-token embedding dataset on load.
- Source text is NVIDIA When2Call revision 0582f77, rendered as the Jev Decision Index renders the test.
Questions about this entry
Short answers drawn from the linked project or page.
It is a Hugging Face dataset from clm-project holding state, action, and hard-negative embeddings for post-training CLM heads. The layout describes 16,510 train pairs and 1,856 validation pairs; the dataset viewer reports 2 rows and 385 MB total file size.
The card gives the command clm-train posttrain --source hf:boost-to7-w2ctrain-when2call-qwen3.6. The dataset is also readable through the Hugging Face Datasets library, with pandas and Polars listed, and auto-converted to Parquet.
Vectors come from Qwen/Qwen3.6-35B-A3B-FP8 with vLLM pooling, last token, L2-normalized, at 2048 dimensions. Metadata records hidden_size 2048, max_model_len 16384, state_head_tokens 2048, negatives_k 3, and multimodal false.
Hard negatives are the other options, at most 8 per row, with near-copies of the gold option dropped. Stored negatives use k=3 and are zero-padded in neg_embeddings.pt, with true counts in neg_counts.pt.
The card states vLLM's pooling of a one-token prompt is noise for this model, so one-token sequences are never embedded. They are zeroed here and 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 license is per-source: CC BY 4.0 for the NVIDIA source and CC BY 4.0 for xLAM calls (Salesforce, via argilla). The vectors are described as a derivative of the source text.
NVIDIA's When2Call training split (sft + pref; the test MCQ is not read), revision 0582f7749df63a96fdc3070932e83e72396ace53, rendered as the index renders the test: tools in the BFCL schema plus question, and the four replies in the index's order.
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 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
arXiv paper testing whether the readout and adaptation choices used for direct selection still hold up in cross-request entailment reranking for JEV decision models. It combines a mean-span readout with selective backbone adaptation and reports accuracy on MC11, an average over eleven reranking configurations.
Jev Models & ReimplementationsArticle#arxiv#research#classification
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.