Long range communication with LoRa has become popular as it avoids the complexity of multi-hop communication at low cost and low energy consumption. LoRa is openly accessible, but its packets are particularly vulnerable to collisions due to long time on air in a shared band. This degrades communication performance. Existing techniques for demodulating LoRa symbols under collisions face challenges such as high computational complexity, reliance on accurate symbol boundary information, or error-prone peak detection methods. In this paper, we introduce CoRa, a symbol detector for demodulating LoRa symbols under severe collisions. CoRa employs a Bayesian classifier to accurately identify the true symbol amidst interference from other LoRa transmissions, leveraging empirically derived features from raw symbol data. Evaluations using realworld and simulated packet traces demonstrate that CoRa clearly outperforms the related state-of-the-art, i.e., up to 29\% better decoding performance than TnB and 178\% better than CIC. Compared to the LoRa baseline demodulator, CoRa magnifies the packet reception rate by up to 11.53x. CoRa offers a significant reduction in computational complexity compared to existing solutions by only adding a constant overhead to the baseline demodulator, while also eliminating the need for peak detection and accurately detecting colliding frames.