diff --git a/src/training/pipeline.py b/src/training/pipeline.py
index cb23715..ec10e67 100644
--- a/src/training/pipeline.py
+++ b/src/training/pipeline.py
@@ -374,8 +374,8 @@ class DataPipeline:
                     split_data[split_name]["X"] = split_df.select(feature_cols)
                     split_data[split_name]["y"] = split_df[label_name]
 
-        # 删除标签为 NaN 的行
-        for split_name in ["train", "val", "test"]:
+        # 删除标签为 NaN 的行（仅在 train/val 上执行，test 集保留用于预测）
+        for split_name in ["train", "val"]:
             if split_name in split_data:
                 y_series = split_data[split_name]["y"]
                 y_nan_count = y_series.null_count()