search

patrik-bartak / atcs-crosslingual-sharing

MIT License
0 stars 0 forks source link

Finetune XLM-R on XNLI EN #8

Closed fanminshi closed 5 months ago

fanminshi commented 5 months ago

The tricky part is to figure out how hugging trainer handle multiple inputs and compute the loss correctly.

patrik-bartak commented 5 months ago

Done

patrik-bartak commented 5 months ago

5 epochs 122720 steps size 16 lr 1e-5 with 10% warmup and linear lr decay

final en validation acc 0.844578313253012

patrik-bartak commented 5 months ago

one shot test zh 0.7337325349301397 hi 0.6872255489021956

one shot validation zh 0.7285140562248996 hi 0.685140562248996 cs 0.7566265060240964 nl ... TBA id ... TBA