Melanoma is the most fatal type of skin cancer. Non-invasive melanoma detection is crucial for preliminary screening and early diagnosis. Among various image based techniques, hyperspectral imaging is a tool with great potential for melanoma detection since it provides highly detailed spectral information beyond the human vision capability. However, so far no hyperspectral image dataset has been published, although some pilot methods have been studied. In this paper, we introduce a hyperspectral dermoscopy image dataset for melanoma detection. This dataset consists of 330 hyperspectral images with 16 spectral bands each in the visible wavelength, containing images of melanoma, dysplastic nevus, and other types, all histopathologically validated. To build a baseline for melanoma detection, we evaluate several classification methods on the dataset.