Video action understanding is a rapidly growing field that has achieved excellent results in various application areas, such as sports and lifestyle applications. However, research that combines computer vision action understanding techniques and the artistic domain of classical ballet choreography is still in its infancy. Publicly available ballet video datasets are limited in number and need more richness to properly explore this specialized field and its extensive collection of actions. Recordings of ballet rehearsals, performances, and competitions have become more readily available on public platforms in recent years, making a substantial amount of data available in this discipline. We propose a novel video dataset, AnnChor, for temporal action localization in ballet choreography. The dataset is notable for its quality and the diversity of ballet actions found in the videos of solo ballet performances. The full dataset comprises 1020 videos with over 25 000 temporal annotations for 11 action classes. We evaluate and provide baseline results for temporal action localization using the Coarse-Fine Network and TriDet models. There is much opportunity to advance computer vision technology to aid the classical dance domain. We hope this dataset will benefit the computer vision community and enable researchers to explore the challenges present in action localization, especially in the context of fine-grained ballet movements. The dataset can be found at https://github.com/dvanderhaar/UJAnnChor .