The multi-task learning setting is different from the original paper

[Aurora-slz]

According to the GLM paper, multi-task learning has two way, one is a mixture of the blank infilling object and the sentence-level objective, another is a mix of the blank infilling object and the document-level objective. But when I read the pre-training code (/config/ds_block_chinese.sh), I found that multi-task learning takes 40% blank infilling object, 30% sentence-level objective, 30% document-level objective. Am I understanding it wrong?

...
gpt_options=" \
       --block-lm \
       --task-mask \
       --bert-prob 0.4 \
       --gap-sentence-prob 0.3 \
...

[duzx16]

The two ways of multi-task learning are designed for ablation study of different objectives. To enable adaption to different downstream tasks, we mix all the three types of objectives in the Chinese model.