GAP-mmmu-pro-vision
收藏资源简介:
该数据集包含多个配置,每个配置对应由Gemma系列模型(如gemma-4-26b-a4b-it和gemma-4-e4b-it)生成的指令跟随数据。数据集中每个样本包含一个多模态指令(问题question和图像images),以及对应的参考答案ref_answer和模型生成的文本generated_texts。数据集旨在用于视觉语言模型的训练或评估,特别是针对指令跟随和思考推理任务。每个样本还记录了输入token长度、生成token长度、预填充token数、完成原因、种子、标签、图像指纹、图像处理器配置、视觉token统计等信息。数据规模方面,每个配置的训练集样本数在50到500之间,总大小从约38MB到434MB不等。配置名称中的“ins-gen”表示指令生成模式,“think-gen”表示思考推理生成模式,视觉token数(vt140/vt280/vt560)表示图像编码后的视觉token数量,最大生成token数(2048/4096)表示文本生成上限。该数据集适用于多模态指令微调、视觉问答、图像描述生成等任务。
This dataset contains multiple configurations, each corresponding to instruction-following data generated by Gemma series models (e.g., gemma-4-26b-a4b-it and gemma-4-e4b-it). Each sample in the dataset includes a multimodal instruction (question and images), along with the corresponding reference answer (ref_answer) and model-generated text (generated_texts). The dataset is designed for training or evaluating vision-language models, particularly for instruction-following and reasoning tasks. Each sample also records information such as input token length, generation token length, prefilled token count, completion reason, seed, label, image fingerprint, image processor configuration, and visual token statistics. In terms of data scale, the number of training samples per configuration ranges from 50 to 500, with total sizes ranging from approximately 38MB to 434MB. In the configuration names, ins-gen indicates instruction generation mode, think-gen indicates reasoning generation mode, the visual token count (vt140/vt280/vt560) represents the number of visual tokens after image encoding, and the maximum generation token count (2048/4096) indicates the upper limit of text generation. This dataset is suitable for tasks such as multimodal instruction fine-tuning, visual question answering, and image caption generation.
数据集详情总结
数据集名称
GAP-mmmu-pro-vision
数据集地址
https://huggingface.co/datasets/elichen-skymizer/GAP-mmmu-pro-vision
数据集概述
该数据集包含多个配置(config),每个配置对应不同的模型生成设置和子采样规模。数据集主要针对视觉问答(VQA)任务,包含问答、图片、模型生成文本及相关元数据。
配置详情
配置列表及其规模
| 配置名称 | 训练集样本数 | 数据集大小(Bytes) | 下载大小(Bytes) |
|---|---|---|---|
| gemma-4-26b-a4b-it-ins-gen-2048-vt280 | 498 | 415,110,812 | 414,160,911 |
| gemma-4-26b-a4b-it-ins-gen-2048-vt280-subsample-50 | 50 | 51,202,321 | 50,073,025 |
| gemma-4-26b-a4b-it-ins-gen-2048-vt280-subsample-100 | 100 | 90,042,001 | 87,773,610 |
| gemma-4-26b-a4b-it-think-gen-4096-vt280 | 493 | 433,979,526 | 430,987,218 |
| gemma-4-26b-a4b-it-think-gen-4096-vt280-subsample-50 | 50 | 42,164,194 | 39,128,500 |
| gemma-4-26b-a4b-it-think-gen-4096-vt280-subsample-100 | 100 | 98,337,544 | 92,408,822 |
| gemma-4-e4b-it-ins-gen-2048-vt140 | 500 | 417,042,336 | 415,987,693 |
| gemma-4-e4b-it-ins-gen-2048-vt140-subsample-50 | 50 | 53,257,276 | 52,265,522 |
| gemma-4-e4b-it-ins-gen-2048-vt140-subsample-100 | 100 | 96,524,053 | 94,355,753 |
| gemma-4-e4b-it-ins-gen-2048-vt280 | 496 | 417,993,830 | 416,816,230 |
| gemma-4-e4b-it-ins-gen-2048-vt280-subsample-50 | 50 | 45,958,100 | 44,714,186 |
| gemma-4-e4b-it-ins-gen-2048-vt280-subsample-100 | 100 | 93,940,761 | 91,212,346 |
| gemma-4-e4b-it-ins-gen-2048-vt560 | 499 | 420,543,048 | 419,338,979 |
| gemma-4-e4b-it-ins-gen-2048-vt560-subsample-50 | 50 | 38,316,919 | 36,828,399 |
| gemma-4-e4b-it-ins-gen-2048-vt560-subsample-100 | 100 | 74,593,773 | 71,488,325 |
| gemma-4-e4b-it-think-gen-4096-vt140 | - | - | - |
(注:最后一个配置的 splits 信息在提供内容中不完整。)
数据字段(所有配置通用)
- source (string):数据来源
- item_id (string):项目ID
- origin_id (int64):原始ID
- sample_seed (int64):采样种子
- question (string):问题文本
- images (list of image):图像列表
- num_images (int32):图像数量
- ref_answer (string):参考答案
- category (string):类别
- input_ids (list of int64):输入token IDs
- input_tokens_len (int64):输入token长度
- generated_texts (string):模型生成的文本
- generated_tokens_len (int64):生成token长度
- n_prefill_tokens (int64):预填充token数
- finish_reason (string):生成结束原因
- seed (int64):随机种子
- labels (list of int64):标签
- image_fingerprint (string):图像指纹
- image_processor_config (string):图像处理器配置
- image_processor_config_hash (string):图像处理器配置哈希
- sum_vision_tokens (int64):视觉token总数
- max_vision_tokens (int64):最大视觉token数
- id (string):唯一标识
- truncated_by_cap (bool):是否被截断
- generation_model_name_or_path (string):生成模型名称或路径
- enable_thinking (bool):是否启用思考模式
主要特征
- 数据集围绕视觉问答(VQA)构建,包含问题、参考回答和模型生成文本。
- 每个样本包含图像数据及图像处理相关信息。
- 配置名称反映了生成模型(如gemma-4-26b-a4b-it、gemma-4-e4b-it)、生成模式(ins/think)、生成长度(2048/4096)以及视觉token数(vt140/vt280/vt560)。
- 提供全量训练集和子采样(50、100样本)版本,便于快速实验。




