BUT-FIT/orca-audio-qa-annotations
收藏资源简介:
ORCA音频问答标注数据集是用于训练和评估ORCA(开放式响应正确性评估)模型的标注数据,ORCA是一个用于音频问答任务的评分模型。该数据集采用三阶段课程训练设计,包括以下配置:1. stage1_pretrain:包含5,332,242个项目,来源于5个LLM法官的合成问答评分;2. stage2_benchmark:包含449,730个项目,来源于5个LLM法官的评估基准评分;3. stage3_mmau_mmar:包含2,447个项目,来源于人类标注者的评估评分;4. stage3_mmau_pro:包含1,240个项目,来源于人类标注者的专业评估评分。数据集总大小在100万到1000万之间,支持文本分类任务,并专注于音频问答、正确性评估、ORCA和评估等标签。
Annotation data for training and evaluating ORCA (Open-ended Response Correctness Assessment), a scoring model for audio question-answering tasks. The dataset is structured with a three-stage curriculum: stage1_pretrain includes 5,332,242 items from 5 LLM judges; stage2_benchmark includes 449,730 items from 5 LLM judges; stage3_mmau_mmar includes 2,447 items from human annotators; stage3_mmau_pro includes 1,240 items from human annotators. It has a size category of 1M<n<10M, supports text-classification tasks, and focuses on tags such as audio-question-answering, correctness-assessment, orca, and evaluation.




