Fix W605 flake8 warning (x5).

examples/contrib/run_openai_gpt.py

@@ -22,8 +22,8 @@
           --model_name openai-gpt \
           --do_train \
           --do_eval \
-          --train_dataset $ROC_STORIES_DIR/cloze_test_val__spring2016\ -\ cloze_test_ALL_val.csv \
-          --eval_dataset $ROC_STORIES_DIR/cloze_test_test__spring2016\ -\ cloze_test_ALL_test.csv \
+          --train_dataset "$ROC_STORIES_DIR/cloze_test_val__spring2016 - cloze_test_ALL_val.csv" \
+          --eval_dataset "$ROC_STORIES_DIR/cloze_test_test__spring2016 - cloze_test_ALL_test.csv" \
           --output_dir ../log \
           --train_batch_size 16 \
 """

transformers/tokenization_xlm.py

@@ -725,10 +725,10 @@ class XLMTokenizer(PreTrainedTokenizer):
             make && make install
             pip install kytea
             ```
-        - [jieba](https://github.com/fxsjy/jieba): Chinese tokenizer *
+        - [jieba](https://github.com/fxsjy/jieba): Chinese tokenizer (*)
             - Install with `pip install jieba`
 
-        \* The original XLM used [Stanford Segmenter](https://nlp.stanford.edu/software/stanford-segmenter-2018-10-16.zip).
+        (*) The original XLM used [Stanford Segmenter](https://nlp.stanford.edu/software/stanford-segmenter-2018-10-16.zip).
         However, the wrapper (`nltk.tokenize.stanford_segmenter`) is slow due to JVM overhead, and it will be deprecated.
         Jieba is a lot faster and pip-installable. Note there is some mismatch with the Stanford Segmenter. It should be fine
         if you fine-tune the model with Chinese supervisionself. If you want the same exact behaviour, use the original XLM