As the digital transformation continues to advance, cross-institutional data sharing has become a crucial means to enhance socio-economic efficiency. However, issues such as data privacy protection, data silos, and compliance severely constrain the practical application of data sharing. In response to this problem, this paper proposes a cross-institutional data sharing mechanism based on federated learning. This mechanism achieves distributed data training through a blockchain and federated learning framework, avoiding direct data transmission and centralized storage, thus enhancing model performance while ensuring data privacy. Additionally, this paper designs an optimized federated learning algorithm that dynamically adjusts weights based on data quality and computational capability, which has been experimentally verified. The experimental results indicate that the proposed mechanism performs well in terms of model convergence speed and accuracy.