fix training: when overlapping chunks in training data, will use head...

fix training: when overlapping chunks in training data, will use head information only from the chunk really taken