---
license: apache-2.0
library_name: transformers
tags:
- generative-embedding
- multimodal
- benchmark
---

# GEB Decoder Checkpoints

<div align="center">
  <a href="https://github.com/LimitedMouse/Generative-Embedding-Benchmark"><img src="https://img.shields.io/badge/GitHub-Code-lightgrey" alt="GitHub Code"></a>
  <a href="https://huggingface.co/datasets/LimitedMouse/Generative-Embedding-Benchmark"><img src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Dataset-blue" alt="Hugging Face Dataset"></a>
</div>

This repository provides the official Qwen decoder checkpoints for the
[Generative Embedding Benchmark](https://github.com/LimitedMouse/Generative-Embedding-Benchmark).
The GitHub repository is the main entry point for installation, training, and
evaluation. Fixed evaluation membership and provenance metadata are hosted in
the [GEB dataset repository](https://huggingface.co/datasets/LimitedMouse/Generative-Embedding-Benchmark).

| Embedder | Mode | Subfolder |
|---|---|---|
| Qwen3-VL-Embedding-2B | Visual-only | `qwen3vl-2b/visual_only` |
| Qwen3-VL-Embedding-2B | VL-joint | `qwen3vl-2b/vl_joint` |
| Qwen3-VL-Embedding-8B | Visual-only | `qwen3vl-8b/visual_only` |
| Qwen3-VL-Embedding-8B | VL-joint | `qwen3vl-8b/vl_joint` |

Each checkpoint contains the trained Qwen3-0.6B readout and its embedding
adapter. The embedding model itself is not bundled; use the matching embedder
and evaluation mode shown above.

```bash
git clone https://github.com/LimitedMouse/Generative-Embedding-Benchmark.git
cd Generative-Embedding-Benchmark
pip install -e '.[eval]'

NPROC=8 BATCH_SIZE=16 bash scripts/evaluate_paper.sh \
  qwen3vl-2b auto test visual_only
```

The checkpoints are intended for reproducing the GEB paper results. The Qwen
models and upstream evaluation datasets remain subject to their own licenses.