Text-based Group Re-Identification (TB G-ReID) is a cross-modal retrieval technology that aims to accurately retrieve specific groups of pedestrians across nonoverlapping camera views using natural language descriptions. Existing text-based person retrieval methods fail to generalize to group re-identification tasks due to their inability to effectively handle issues such as interference from unreliable matching pairs, one-to-many mismatching, and more critically, challenges posed by variations in group size and layout in group scenarios. To address these, we propose the Uncertainty Relation Reasoning and Minimum Permutation Distance for Text-based Group Re-Identification (URMP), which comprises two core components: 1) the Uncertainty Relation Reasoning (URR) module, which leverages evidence theory and the Fisher Information Matrix to quantify and mitigate cross-modal uncertainty, enhancing the robustness of text-image feature alignment; 2) We design a matching module based on the Minimum Permutation Distance (MPD) method, which introduces permutation-invariant metrics to calculate group similarity and resolves layout sensitivity by generating order-agnostic group representations. To support TB G-ReID research, we constructed the CMG-PEDES dataset, which contains 4,052 image-text pairs of 427 groups (involving 1,013 identities). Extensive experiments on the CMG-PEDES dataset demonstrate the effectiveness of the proposed model.