--- dataset_info: - config_name: A-OKVQA features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 14048199 num_examples: 1000 download_size: 1168340 dataset_size: 14048199 - config_name: CIFAR-100 features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 1519890 num_examples: 1000 download_size: 20544 dataset_size: 1519890 - config_name: CIRR features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 70162098 num_examples: 1000 download_size: 1565489 dataset_size: 70162098 - config_name: ChartQA features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 14354641 num_examples: 1000 download_size: 1434448 dataset_size: 14354641 - config_name: Country211 features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 3678000 num_examples: 1000 download_size: 31556 dataset_size: 3678000 - config_name: DocVQA features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 23044459 num_examples: 1000 download_size: 1734476 dataset_size: 23044459 - config_name: EDIS features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 184208708 num_examples: 1000 download_size: 3350382 dataset_size: 184208708 - config_name: FashionIQ features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 71169665 num_examples: 1000 download_size: 1729457 dataset_size: 71169665 - config_name: GQA features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 40809641 num_examples: 1000 download_size: 1764457 dataset_size: 40809641 - config_name: HatefulMemes features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 184890 num_examples: 1000 download_size: 9972 dataset_size: 184890 - config_name: ImageNet-1K features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 28773890 num_examples: 1000 download_size: 185019 dataset_size: 28773890 - config_name: ImageNet-A features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 28772890 num_examples: 1000 download_size: 147780 dataset_size: 28772890 - config_name: ImageNet-R features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 3456890 num_examples: 1000 download_size: 23656 dataset_size: 3456890 - config_name: InfographicsVQA features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 19114439 num_examples: 1000 download_size: 1439837 dataset_size: 19114439 - config_name: MSCOCO features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 97759085 num_examples: 1000 download_size: 1681753 dataset_size: 97759085 - config_name: MSCOCO_i2t features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 60201740 num_examples: 1000 download_size: 1785583 dataset_size: 60201740 - config_name: MSCOCO_t2i features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 87127008 num_examples: 1000 download_size: 1296167 dataset_size: 87127008 - config_name: N24News features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 630658 num_examples: 1000 download_size: 110698 dataset_size: 630658 - config_name: NIGHTS features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 75116000 num_examples: 1000 download_size: 1528646 dataset_size: 75116000 - config_name: OK-VQA features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 15332578 num_examples: 1000 download_size: 1564823 dataset_size: 15332578 - config_name: OVEN features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 717934263 num_examples: 1000 download_size: 406792141 dataset_size: 717934263 - config_name: ObjectNet features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 2036000 num_examples: 1000 download_size: 27132 dataset_size: 2036000 - config_name: Place365 features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 7045000 num_examples: 1000 download_size: 89866 dataset_size: 7045000 - config_name: RefCOCO features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 96493941 num_examples: 1000 download_size: 1858145 dataset_size: 96493941 - config_name: RefCOCO-Matching features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 145712476 num_examples: 1000 download_size: 2879385 dataset_size: 145712476 - config_name: SUN397 features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 7990000 num_examples: 1000 download_size: 118447 dataset_size: 7990000 - config_name: ScienceQA features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 23870406 num_examples: 1000 download_size: 958782 dataset_size: 23870406 - config_name: TextVQA features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 17435986 num_examples: 1000 download_size: 1571656 dataset_size: 17435986 - config_name: VOC2007 features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 368000 num_examples: 1000 download_size: 13813 dataset_size: 368000 - config_name: VisDial features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 67989850 num_examples: 1000 download_size: 1730820 dataset_size: 67989850 - config_name: Visual7W features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 22047066 num_examples: 1000 download_size: 1564788 dataset_size: 22047066 - config_name: Visual7W-Pointing features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 94906832 num_examples: 1000 download_size: 1299380 dataset_size: 94906832 - config_name: VisualNews_i2t features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 118329649 num_examples: 1000 download_size: 81491360 dataset_size: 118329649 - config_name: VisualNews_t2i features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 97176206 num_examples: 1000 download_size: 1763677 dataset_size: 97176206 - config_name: VizWiz features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 20550246 num_examples: 1000 download_size: 1425789 dataset_size: 20550246 - config_name: WebQA features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 197701404 num_examples: 1000 download_size: 3257136 dataset_size: 197701404 - config_name: Wiki-SS-NQ features: - name: qry_text dtype: string - name: qry_img_path dtype: string - name: tgt_text sequence: string - name: tgt_img_path sequence: string splits: - name: test num_bytes: 74583207 num_examples: 1000 download_size: 1900579 dataset_size: 74583207 configs: - config_name: A-OKVQA data_files: - split: test path: A-OKVQA/test-* - config_name: CIFAR-100 data_files: - split: test path: CIFAR-100/test-* - config_name: CIRR data_files: - split: test path: CIRR/test-* - config_name: ChartQA data_files: - split: test path: ChartQA/test-* - config_name: Country211 data_files: - split: test path: Country211/test-* - config_name: DocVQA data_files: - split: test path: DocVQA/test-* - config_name: EDIS data_files: - split: test path: EDIS/test-* - config_name: FashionIQ data_files: - split: test path: FashionIQ/test-* - config_name: GQA data_files: - split: test path: GQA/test-* - config_name: HatefulMemes data_files: - split: test path: HatefulMemes/test-* - config_name: ImageNet-1K data_files: - split: test path: ImageNet-1K/test-* - config_name: ImageNet-A data_files: - split: test path: ImageNet-A/test-* - config_name: ImageNet-R data_files: - split: test path: ImageNet-R/test-* - config_name: InfographicsVQA data_files: - split: test path: InfographicsVQA/test-* - config_name: MSCOCO data_files: - split: test path: MSCOCO/test-* - config_name: MSCOCO_i2t data_files: - split: test path: MSCOCO_i2t/test-* - config_name: MSCOCO_t2i data_files: - split: test path: MSCOCO_t2i/test-* - config_name: N24News data_files: - split: test path: N24News/test-* - config_name: NIGHTS data_files: - split: test path: NIGHTS/test-* - config_name: OK-VQA data_files: - split: test path: OK-VQA/test-* - config_name: OVEN data_files: - split: test path: OVEN/test-* - config_name: ObjectNet data_files: - split: test path: ObjectNet/test-* - config_name: Place365 data_files: - split: test path: Place365/test-* - config_name: RefCOCO data_files: - split: test path: RefCOCO/test-* - config_name: RefCOCO-Matching data_files: - split: test path: RefCOCO-Matching/test-* - config_name: SUN397 data_files: - split: test path: SUN397/test-* - config_name: ScienceQA data_files: - split: test path: ScienceQA/test-* - config_name: TextVQA data_files: - split: test path: TextVQA/test-* - config_name: VOC2007 data_files: - split: test path: VOC2007/test-* - config_name: VisDial data_files: - split: test path: VisDial/test-* - config_name: Visual7W data_files: - split: test path: Visual7W/test-* - config_name: Visual7W-Pointing data_files: - split: test path: Visual7W-Pointing/test-* - config_name: VisualNews_i2t data_files: - split: test path: VisualNews_i2t/test-* - config_name: VisualNews_t2i data_files: - split: test path: VisualNews_t2i/test-* - config_name: VizWiz data_files: - split: test path: VizWiz/test-* - config_name: WebQA data_files: - split: test path: WebQA/test-* - config_name: Wiki-SS-NQ data_files: - split: test path: Wiki-SS-NQ/test-* --- Embedding models have been crucial in enabling various downstream tasks such as semantic similarity, information retrieval, and clustering. Recently, there has been a surge of interest in developing universal text embedding models that can generalize across tasks (e.g., MTEB). However, progress in learning universal multimodal embedding models has been relatively slow despite their importance. In this work, we aim to explore the potential for building universal embeddings capable of handling a wide range of downstream tasks. Our contributions are twofold: (1) MMEB (Massive Multimodal Embedding Benchmark), which covers 4 meta-tasks including classification, question answering, retrieval, and visual grounding and 36 datasets, including 20 training and 16 evaluation datasets, and (2) VLM2Vec (Vision-Language Model => Vector), a contrastive training framework that converts any state-of-the-art vision-language model into an embedding model via training on MMEB. Unlike previous models such as CLIP and BLIP, VLM2Vec can process any combination of images and text to generate a fixed-dimensional vector based on task instructions. We build a series of VLM2Vec models on Phi-3.5-V and evaluate them on MMEB's evaluation split. Our results show that VLM2Vec achieves an absolute average improvement of 10% to 20% over existing multimodal embedding models on both in-distribution and out-of-distribution datasets in MMEB.