Fair exchange of trained machine learning (ML) model allows data owners with limited time, resources or expertise to purchase the well-trained ML models from experts. For fewer computation, current works employ sampling methods to provide workload assurance (i.e., the claimed resources consumption of the seller indeed corresponds to the actual workload), but these methods may fail to detect the seller’s cheat with a non-negligible probability. To address this problem, we propose fair exchange framework that could bring sound and complete workload assurance by on-chain reduplicating and quantificating the training process. Meanwhile, we use state-of-the-art zero-knowledge proof algorithm to guarantee service correctness of the training (i.e., the trained ML models provided by the seller meets the requirements of the buyer) and deploy higher-speed consortium blockchain as underlying platform to accelerate the on-chain execution. We evaluate our framework with six most used machine learning algorithms, and develop corresponding optimizations to construct circuits that enables efficient proof generation and verification. The evaluation results show the practical performance of our framework.