Rouge very low using provided example

[yjqiu]

Hi, thanks for the demo script and nice tutorial. I tried to use the example code included in summarization.py as below. But I only get Rouge ~0.1 for testing data. Is there anything I didn't notice? Or I have to use all training data to get Rouge 43 as shown in the slide. Thanks!

PYTHONWARNINGS="ignore" CUDA_VISIBLE_DEVICES=6,7   python summarization.py  \
    --fp16  --batch_size 2  --grad_accum 1 --grad_ckpt   \
    --max_input_len  16384 --attention_window  1024