crystina-z/mdpr-adapter.w-mbert.pft-msmarco

Tevatron

bs=512
epoch=40

save_steps=4000 
backbone=bert-base-multilingual-cased
output_dir=mlm.bs-$bs.epoch-$epoch.$backbone

WANDB_PROJECT=mlm-mrtydi-DDR \
    python examples/dense-adapter/dense-adapter-train.py \
    --output_dir $output_dir \
    --model_name_or_path $backbone \
    --tokenizer_name bert-base-multilingual-cased \
    --save_steps $save_steps \
    --dataset_name Tevatron/msmarco-passage \
    --fp16 \
    --per_device_train_batch_size $bs \
    --train_n_passages 2 \
    --learning_rate 1e-5 \
    --q_max_len 32 \
    --p_max_len 128 \
    --num_train_epochs $epoch \
    --logging_steps 100 \
    --overwrite_output_dir \
    --dataloader_num_workers 4 \