The Extractive Question Answering (EQA) is a fundamental task in Natural Language Processing domain, focusing on the identification of a correct answer span (a sequence of continuous words) over the given question and passage. Recent years have witnessed a dramatically increasing interest in harnessing the Large Language Models (LLMs) for EQA. Yet, the existence of missing labeled training data in many EQA datasets has a negative impact on model training. These missing labeled instances can be attributed to inaccuracies in label annotations, format inconsistencies, and etc. Consequently, this paper introduces a novel training algorithm, termed Adaptive Denoising Training (ADT) to adaptively manipulate missing labeled instances during the training process. Specifically, the proposed ADT is characterized by an adaptive loss function, to dynamically adjust the contribution (or weights) of missing labeled data in each iteration, thereby minimizing their impact on model training. Extensive experiments over three benchmarks demonstrate that the proposed ADT algorithm achieves notable improvement compared to conventional training methods. Additionally, ADT also achieves strong performance in low-resource settings.