existence of training set seems to affect deploy result

[jacksampeng]

In the doc it says: To perform NER on some plain texts using a pre-trained model:

python3.5 main.py --train_model=False --use_pretrained_model=True --dataset_text_folder=../data/example_unannotated_texts --pretrained_model_folder=../trained_models/conll_2003_en 

In the beginning I just created a deploy folder alongside the train, test and valid folders and put some unannotated text in it. I then ran this command. It took a while to run. So I created another folder and just added deploy folder. To my surprise, the results on the deploy set is very different, with the first method (with training data) doing much better. I tried to find the reason in the code but so far haven't found anything. Does anyone know?