lance-format/vqav2-lance

VQAv2（Lance格式）是VQAv2数据集的Lance格式版本，基于COCO图像进行开放式视觉问答。数据集每行包含一个（图像、问题、10个标注者答案）三元组，其中图像以内联JPEG字节形式存储，问题为自然语言文本。此外，每行还提供配对的CLIP图像嵌入和问题嵌入（使用OpenCLIP ViT-B/32模型生成，512维，余弦归一化），这些嵌入共享同一空间，支持跨模态检索。数据集包括VQAv2的问题类型和答案类型分类，以及共识答案（multiple_choice_answer）。预构建的索引覆盖嵌入向量、问题文本、答案分类和ID，支持快速搜索和筛选。数据集仅包含验证集分割（214,354行），训练集可通过其他方式获取。数据格式优化了随机访问、多模态支持和高效查询，适用于机器学习训练、检索和数据分析。VQAv2 (Lance Format) is a Lance-formatted version of the VQAv2 dataset, featuring open-ended visual question answering on COCO images. Each row consists of an (image, question, 10 annotator answers) triple, with images stored as inline JPEG bytes and questions as natural language text. Additionally, each row includes paired CLIP image and question embeddings (generated using the OpenCLIP ViT-B/32 model, 512-dimensional, cosine-normalized) that share the same embedding space, enabling cross-modal retrieval. The dataset incorporates VQAv2s question_type and answer_type taxonomy, along with the consensus multiple_choice_answer. Pre-built indices cover embeddings, question text, answer taxonomy, and IDs, facilitating fast search and filtering. Only the validation split is included (214,354 rows), with training data available through alternative sources. The format is optimized for random access, multimodal support, and efficient querying, making it suitable for machine learning training, retrieval, and data analysis.