Jev Boost to7 ToolRetTrain ToolRet Qwen3.6-35B-A3B-FP8 Embeddings

It supplies state, action, and hard-negative embeddings for post-training CLM heads, reshaped from mangopy/ToolRet-Training-20w into decisions that follow the Jev Decision Index's Tool Use & Agentic Decisions request shape. It is consumed through clm_train.data.jev_boost, documented in docs/jev-boost.md.

Vectors are pooled from Qwen/Qwen3.6-35B-A3B-FP8 using vLLM pooling on the last token, L2-normalized, at hidden size 2,048. Metadata records max_model_len 16,384, state_head_tokens 2,048, renderer clm_train.render.Renderer, multimodal false, and negatives_k 1.

State is the record's text as the index renders the target benchmark's request, followed by that benchmark's own instructions; states never list the options. Action is the gold option's text. Hard negatives are the other options, at most 8 per row, with near-copies of the gold dropped.

The dataset is loaded by clm_train, with the source referenced as hf:boost-to7-toolrettrain-toolret-qwen3.6-35b-a3b-fp8-embeddings for post-training. Files follow the split-per-directory layout that clm_train expects: per-split .pt tensors plus metadata.json listing one dict per row.

One-token sequences are never embedded, because vLLM's pooling of a one-token prompt is noise for this model. They are stored as zeros, 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 card states no licence of its own and defers to the source's terms: ToolRet code is Apache-2.0, while tools from ToolBench, APIGen, and ToolACE keep their own terms. The vectors are a derivative of the source text, so those terms follow the derivative.

Items whose state shares a paragraph with a Jev suite item are dropped, and the set is capped at 20,000 rows by hash sample. Negative embeddings are zero-padded with counts in neg_counts.pt, and train carries 18,009 pairs versus 1,991 in validation.