lance-format/textvqa-lance-colab
收藏资源简介:
TextVQA Lance Colab 是一个小型 Lance 格式的数据集,来源于 TextVQA,用于 TMLS 2026 研讨会演示和在子集上微调视觉语言模型。该数据集包含 1,000 行数据,其中 600 行用于训练,400 行用于验证,每行为一个视觉问答示例,涉及包含场景文本的图像。数据集包括内联图像字节、自然语言问题、10 个参考答案、OCR 令牌、图像类别标签以及成对的 512 维图像/问题嵌入。训练分割还包括预计算的固定形状张量,适用于 Colab 微调管道。数据集支持通过 Hugging Face 的 datasets 库、LanceDB 或 Lance 加载,适用于多模态检索、模型训练和演示。
TextVQA Lance Colab is a small Lance-formatted subset of TextVQA, sourced from lmms-lab/textvqa, for the TMLS 2026 Workshop Demo and fine-tuning a VLM on the subset. It contains 1,000 rows total, with 600 train examples and 400 validation examples. Each row is one visual question-answering example over an image that contains scene text, including inline image bytes, a natural-language question, 10 reference answers, OCR tokens, image-class labels, and paired 512-dimensional image/question embeddings. The train split also includes precomputed fixed-shape tensors for a Colab fine-tuning pipeline. The dataset can be loaded via Hugging Face datasets, LanceDB, or Lance, and is suitable for multimodal retrieval, model training, and demonstrations.




