pretty_name: Generative Embedding Benchmark
license: apache-2.0
task_categories:
- visual-question-answering
Generative Embedding Benchmark
This repository publishes the fixed membership and provenance metadata for the Generative Embedding Benchmark (GEB). GEB contains 1,800 development questions and 900 held-out test questions spanning natural images, scene text, and visual documents.
GEB evaluates how much answer-relevant visual information a dense embedding makes accessible to a generative decoder. The GitHub repository is the main entry point for installation, training, and evaluation. Official Qwen3-VL-Embedding-2B/8B decoder weights are provided in the checkpoint repository.
This dataset repository does not redistribute source images, questions,
answers, or decoder training data. GEB task definitions load the eight original
public evaluation datasets through lmms-eval and filter them with dev.json
or test.json. The source datasets remain subject to their own licenses and
access requirements.
Files
dev.json,test.json: compact source-row membership used at runtime.manifests/dev/,manifests/test/: categories, provenance identifiers, image hashes, and image-group identifiers used for auditing.sources.json: pinned upstream dataset revisions.summary.json: sampling quotas and aggregate attrition statistics.leakage_audit.json: development/test image-leakage audit.
Citation
@misc{li2026generativeembeddingbenchmark,
title={Generative Embedding Benchmark: How Much Information Survives in a Dense Embedding?},
author={Yun Li and Biao Yang and Peixi Wu and Yunhao Zhou and Mingzhou Jiang and Wei Yuan and Fan Yang and Wenwu Ou},
year={2026},
eprint={2608.06972},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2608.06972},
}
The Apache-2.0 license applies only to GEB-authored manifests and metadata. It does not alter the terms of the upstream datasets.