Automatic speech recognition (ASR) systems have become widespread and have broad categories of usage and users. It is therefore crucial for ASR models to be robust to speaker variations, particularly accents. Since ASR models perform best on their domain of training, researchers try to improve them by gathering large quantities of accented speech data. However, the availability of such data is often very limited, especially for languages other than English. In this work, we introduce a new corpus of Quebec-accented French that we call the CEREALES dataset. It is a large, gender-balanced corpus of more than 600 hours of spontaneous speech collected during a commission of inquiry in Quebec. We clean the transcripts and align them with the audio, making the dataset useful for both pre-training and fine-tuning models. We share it for the benefit of the research community. In addition, we run ASR experiments and show how this dataset can help address the challenges of accented ASR.