jiminsun / multidim-evaluation

char-encoder

Environment setup

conda env create -f requirements.yml -n $ENV_NAME

Training

Single language training

cd xtreme
bash in_language_finetuning.sh $MODEL $TASK $GPU

• $MODEL is either bert-multilingual-base-cased, google/canine-s, google/canine-c, google/mt5-small, or google/byt5-small
• For larger t5 style models, change small to base or large
• $TASK is either tydiqa, xnli, or panx

Multilingual training (multitasking)

mBERT, CANINE models

cd xtreme 
# tydiqa
bash scripts/train_tydiqa_all.sh $MODEL $GPU

# xnli
bash scripts/train_xnli.sh $MODEL $GPU

# panx
bash scripts/train_panx.sh $MODEL $GPU $LANGS
# LANGS: "en,hi,bn,ur,sw,ar,de,el,es,fi,fr,ko,ru,te,zh,tr,ta,id"