CHAI_testset
收藏资源简介:
CHAI(基于批判的人机协同)数据集是一个用于评估精确视频描述监督框架的评估数据集。该框架的核心是让训练有素的人类专家与模型生成的初始描述(称为预描述)协同工作,专家提供修正性的批判意见,从而引导生成改进后的最终描述(称为后描述)。数据集基于一个结构化的描述规范构建,该规范涵盖了视频中的主体、场景、动作、空间布局和摄像机动态等多个维度,其视觉基元由专业电影制作人参与开发。数据集包含多个JSON格式的评估文件,分别用于不同的评估任务:`test_split.json`包含原始评估数据,每条数据包括视频路径、模型生成的预描述、人工撰写的批判、修订后的最终描述(后描述)及相关元数据;`eval_caption_generation_test.json`格式化为描述生成任务,将视频与任务指令及最终描述配对;`eval_critique_generation_test.json`格式化为批判生成任务,将视频、任务指令及待批判的描述配对;`eval_caption_yes_or_no_test.json`格式化为奖励建模(二元对齐评分)任务。该数据集适用于视频到文本的多任务评估,特别是视频描述生成、描述批判生成以及描述质量的对齐评分等研究场景。
The CHAI (Critique-based Human-AI Collaboration) dataset is an evaluation dataset designed to assess a precise video description supervision framework. The core of this framework involves trained human experts collaborating with model-generated initial descriptions (pre-descriptions). Experts provide corrective critiques to guide the generation of improved final descriptions (post-descriptions). The dataset is built on a structured description specification that covers multiple dimensions of videos, including subjects, scenes, actions, spatial layouts, and camera dynamics, with visual primitives developed with input from professional filmmakers. The dataset includes multiple JSON format evaluation files for different tasks: `test_split.json` contains raw evaluation data, each entry comprising video paths, model-generated pre-descriptions, human-written critiques, revised final descriptions (post-descriptions), and related metadata; `eval_caption_generation_test.json` is formatted for description generation tasks, pairing videos with task instructions and final descriptions; `eval_critique_generation_test.json` is formatted for critique generation tasks, pairing videos, task instructions, and descriptions to be critiqued; `eval_caption_yes_or_no_test.json` is formatted for reward modeling (binary alignment scoring) tasks. This dataset is suitable for multi-task evaluation in video-to-text scenarios, particularly for research in video description generation, description critique generation, and alignment scoring of description quality.
数据集概述
CHAI_testset 是一个用于评估视频字幕生成和批评生成的数据集,基于 CHAI (Critique-based Human–AI) 框架构建,旨在通过人类与人工智能协作实现精确的视频语言描述。
数据集任务类别
- video-text-to-text: 视频到文本的转换任务(视频字幕生成与评价)。
数据集背景
该数据集是为 CHAI 评估系统设计的测试集。CHAI 是一个监督框架,由受过训练的人类专家与模型生成的预字幕配对,专家提供修正性批评,指导模型将预字幕改进为更优的后字幕。数据集基于结构化字幕规范构建,涵盖主体、场景、运动、空间布局和摄像机动态,并借鉴了专业电影制作人的视觉基元。
数据集文件结构
数据集包含以下评估文件:
test_split.json: 原始评估数据。每条记录包含视频路径、模型生成的预字幕、人类撰写的批评、修订后的最终字幕(后字幕)及相关元数据。eval_caption_generation_test.json: 为字幕生成任务格式化。将视频与任务指令及最终后字幕配对。eval_critique_generation_test.json: 为批评生成任务格式化。将视频、任务指令及待批评的字幕配对。eval_caption_yes_or_no_test.json: 为奖励建模(二元对齐评分)任务格式化。
数据集使用示例
可使用 Hugging Face CLI 下载完整数据集(包含视频和评估 JSON 文件):
bash pip install huggingface_hub hf download datasets/chancharikm/CHAI_testset --local-dir ./CHAI_testset
引用
如果您使用了本数据集,请引用以下论文:
@inproceedings{chai2026, title = {Building a Precise Video Language with Human--AI Oversight}, author = {Zhiqiu Lin and Chancharik Mitra and Siyuan Cen and Isaac Li and Yuhan Huang and Yu Tong Tiffany Ling and Hewei Wang and Irene Pi and Shihang Zhu and Ryan Rao and George Liu and Jiaxi Li and Ruojin Li and Yili Han and Yilun Du and Deva Ramanan}, booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}, year = {2026} }




