AndyShao90/ReefSet-C

---
license: apache-2.0
language:
- en
pretty_name: RS-C
size_categories:
- 100K<n<1M
---

ReefSet-C (RS-C) serves as a benchmark for [DHAuDS](https://arxiv.org/abs/2511.18421) in the context of test-time adaptation for audio classification.
RS-C, a subset of [ReefSet](https://zenodo.org/records/11071202), utilizes only test-set samples to address domain shift. Each sample in RS-C has a duration
of 1.88 seconds, a sample rate of 16 kHz, and belongs to one of 37 classes. The dataset comprises seven corruption sets: WHN, ENQ, END1, END2, ENSC, TST,
and PSH. Each corruption set is divided into two levels, L1 and L2, where L2 indicates a higher degree of complexity. RS-C is further divided into
evaluation and adaptation sets, each comprising 239,918 samples and including both levels of all seven corruptions. In total, RS-C consists of 479,836 samples,
with each corruption level, such as WHN-L1 or WHN-L2, containing 17,137 samples.

许可证：Apache-2.0
语言：英语
展示名称：RS-C
样本规模类别：100K < 样本数 < 1M

ReefSet-C（下称RS-C）是面向音频分类测试时适配场景的DHAuDS基准数据集，相关DHAuDS详细信息可参考论文链接：https://arxiv.org/abs/2511.18421。RS-C为ReefSet数据集的子集（ReefSet数据集链接：https://zenodo.org/records/11071202），仅采用测试集样本以应对域偏移问题。该数据集的每个样本时长为1.88秒，采样率为16 kHz，隶属于37个类别之一。数据集包含7类扰动集：WHN、ENQ、END1、END2、ENSC、TST及PSH。每类扰动集均划分为L1与L2两个强度等级，其中L2代表更高的复杂度水平。RS-C进一步被划分为评估集与适配集，二者各包含239,918个样本，且涵盖全部7类扰动的两个强度等级。RS-C总计包含479,836个样本，其中每个扰动等级（如WHN-L1或WHN-L2）均包含17,137个样本。