diff --git a/paddle/fluid/framework/data_feed.cc b/paddle/fluid/framework/data_feed.cc
index 5fb141f3c1e4e58dbae19a846fbbb8a0be958b4d..291d8ffc3c3334c2836e1651a8997984bba084e1 100644
--- a/paddle/fluid/framework/data_feed.cc
+++ b/paddle/fluid/framework/data_feed.cc
@@ -259,6 +259,14 @@ bool MultiSlotDataFeed::CheckFile(const char* filename) {
         return false;
       }
     }
+    // It may be added '\t' character to the end of the output of reduce
+    // task when processes data by Hadoop(when the output of the reduce
+    // task of Hadoop has only one field, it will add a '\t' at the end
+    // of the line by default, and you can use this option to avoid it:
+    // `-D mapred.textoutputformat.ignoreseparator=true`), which does
+    // not affect the correctness of the data. Therefore, it should be
+    // judged that the data is not normal when the end of each line of
+    // data contains characters which are not spaces.
     while (endptr - str != len) {
       if (!isspace(*(endptr++))) {
         VLOG(0)