用unimol+模型 inference出的pos_pred取同一个id的均值吗？

在make_pcq_test_dev_submission.py文件中gap_pred是取同一个id对应值的平均，大多数id是8个值的平均。然而在预测出的pos_pred(原子坐标信息中)，为什么相同id得到的pos_pred形状会不一样？也就是说同一个分子的原子数量对不上？如下是某个id预测出的pos_pred
形状为Shape mismatch for id group: [(19, 3), (19, 3), (19, 3), (19, 3), (19, 3), (19, 3), (19, 3), (15, 3)]，第8个预测出来的pos_pred只有15个原子对应的坐标信息，而其它为19个原子。
想请问这是什么情况？以及想得到某个id的最终pos_pred值，是采用什么方法？除去有异常原子数量的预测值，然后其余取均值吗？

Start by reading make_pcq_test_dev_submission.py, especially the grouping and averaging logic for gap_pred and pos_pred. Trace why entries with the same id can have different coordinate shapes, then document the cause and establish the correct rule for producing a final pos_pred when atom counts differ.

Written by the indexing model from the issue text.