Change to the README file to add SWAG results.

README.md

@@ -441,13 +441,25 @@ python run_swag.py \
   --do_train \
   --do_eval \
   --data_dir $SWAG_DIR/data
-  --train_batch_size 10 \
+  --train_batch_size 4 \
   --learning_rate 2e-5 \
   --num_train_epochs 3.0 \
   --max_seq_length 80 \
   --output_dir /tmp/swag_output/
 ```
 
+Training with the previous hyper-parameters gave us the following results:
+```
+eval_accuracy = 0.7776167149855043
+eval_loss = 1.006812262735175
+global_step = 55161
+loss = 0.282251750624779
+```
+
+The difference with the `81.6%` accuracy announced in the Bert article
+is probably due to the different `training_batch_size` (here 4 and 16
+in the article).
+
 ## Fine-tuning BERT-large on GPUs
 
 The options we list above allow to fine-tune BERT-large rather easily on GPU(s) instead of the TPU used by the original implementation.