---
license: apache-2.0
base_model:
- Qwen/Qwen3-VL-Embedding-2B
pipeline_tag: feature-extraction
tags:
- embeddings
- pytorch
---

# About

This repository contains precalculated **Feature Extraction** matrices (embeddings) designed for **Sigma-Captioner**. These caches allow for high-speed image-to-text alignment and similarity searching without re-encoding the entire vocabulary on every run.

# Compatible with any project, but recommended for use with https://github.com/uninterruptedpowersupply3-NEW/Sigma-Captioner/tree/SGLANG

### 📊 Dataset Sources
*   **English:** [dwyl/english-words](https://github.com/dwyl/english-words)
*   **Anime/Visual:** [cagliostrolab/860k-ordered-tags-json](https://huggingface.co/datasets/cagliostrolab/860k-ordered-tags-json) (Danbooru tags)

### 📂 Directory Structure & Usage
The embeddings are split into two versions based on reliability:

| Path | Contents | Recommendation |
| :--- | :--- | :--- |
| `root/` | `vocab_hybrid_matrix.pt` | **Base Use:** Stable and recommended for general tasks. |
| `DANANDENG/` | `vocab_hybrid_meta.pt` | **Experimental:** Hybrid Anime + English merge. |

> [!WARNING]  
> **Experimental Warning:** The Anime tags in the `DANANDENG` directory are prone to hallucinations regarding specific characters. Use the root files for production/standard tagging and reserve the `DANANDENG` files for experimental research.

*   **Model:** Qwen/Qwen3-VL-Embedding-2B
*   **Format:** PyTorch Tensor (.pt) + JSON Metadata