AbstractPhil/cc-task1-json
收藏资源简介:
该数据集将Conceptual Captions(CC captions)转换为结构化JSON格式,专门用于任务1(task_1)。通过使用Qwen3.5-0.8B LoRA适配器模型,对原始英文图像描述进行主观分析,提取并转换为包含类似评估的JSON数据。数据集包含有效性评分(如schema_valid和grounded),其中grounded为true的行可用于训练。数据格式包括id、原始描述、args字段(含subjects、actions、setting等结构化信息),适用于文本生成任务,并支持工具使用和结构化输出。数据集规模在1M到10M之间,语言为英语,配置多个分片(shard_0、shard_1、shard_2)。
Conceptual Captions converted to literal-extraction JSON via the Qwen3.5-0.8B LoRA adapter for task_1, with schema validity and grounding scores. The dataset transforms plain English image prompts into structured JSON containing subjective assessments, including fields like subjects, actions, and setting. It is designed for text-generation tasks, with tags such as conceptual-captions, tool-use, and structured-output. Rows with grounded==true are training-ready, and the dataset is sharded for efficient handling.



