The quality assurance of software containing learning components has become a critical issue in software engineering. Dataset, akin to oil in machines, serves as the core asset for this type of software, directly determining its performance and reliability. Similar to conventional software driven by vast amounts of code, where bugs in the code can severely impair its performance, for intelligent software driven by data, bugs in the data can also impair their performance. In this paper, we propose IDEX, a method to detect and repair semantically mismatched pairs within text dataset. IDEX is powered by LLM with powerful semantic understanding capabilities and incorporates an iterative feedback mechanism. Based on well-designed strategies, IDEX can proactively learn the defect features identified by the LLM and prioritize more potential defects for the LLM to repair. Moreover, IDEX progressively learns the features of defects during iteration and eventually performs defect prediction on data outside the LLM token budget. Through extensive experiments, we have demonstrated the effectiveness of IDEX and its individual modules, as well as the capability of using IDEX to imporve the performance of state-of-the-art text generation models. Additionally, when applied to a widely-used real-world text dataset, IDEX can report and repair twice the number of defects as the direct use of LLMs under the same token expenditure.