hltcoe/RankVideo-Dataset

RANKVIDEO是一个基于推理的视频检索重排序器，明确使用视频内容对查询-视频对进行推理以评估相关性。该数据集主要用于训练和评估RANKVIDEO，基于MultiVENT 2.0基准。数据集包含训练数据、视频到查询的映射以及第一阶段检索结果。数据格式为JSONL，每个训练样本包括查询ID、查询文本、视频ID、视频路径、真实标签、教师概率、证据（字幕和ASR）以及教师推理痕迹。

RANKVIDEO is a reasoning-based reranker for video retrieval that explicitly reasons over query-video pairs using video content to assess relevance. This dataset is primarily based on the MultiVENT 2.0 benchmark and includes training data, mappings of video IDs to candidate query IDs, and first-stage retrieval results. The data format is JSONL, with each training sample containing query ID, query text, video ID, video paths, true label, teacher probability, evidence (caption and ASR), and teacher reasoning traces.