In this paper, we propose a novel solution combining federated learning and blockchain within the Kafka-ML framework to improve model training reliability using data streams. Federated learning enables decentralized training across multiple data sources while preserving privacy, and blockchain ensures the integrity and security of the training process. Kafka-ML is an open-source framework designed to manage ML/AI pipelines with data streams, and its integration with federated learning and blockchain provides a robust solution for scalable and secure machine learning in distributed environments. The proposed approach is evaluated using a large-scale real-world use case in an industrial IoT setting, demonstrating its potential for reliable model training with continuous data streams.