crystina-z/marginmse.teacher-monoXLMR.pft-msmarco.epoch-5.32x2.lr.7e-6.pft-msmarco

Using tevatron, unpushed code

bs=32
lr=7e-6

gradient_accumulation_steps=1
real_bs=$(($bs / $gradient_accumulation_steps))
echo "real_bs: $real_bs"
echo "expected_bs: $bs"
sleep 1s

epoch=5
teacher=crystina-z/monoXLMR.pft-msmarco

dataset=Tevatron/msmarco-passage && dataset_name=enMarco
output_dir=margin-mse.distill/teacher-$(basename $teacher).student-mbert.epoch-${epoch}.${bs}x2.lr.$lr.data-$dataset_name.$commit_id
mkdir -p $output_dir

CUDA_VISIBLE_DEVICES=$device WANDB_PROJECT=distill \
python examples/distill_marginmse/distil_train.py \
  --output_dir $output_dir \
  --model_name_or_path bert-base-multilingual-cased \
  --teacher_model_name_or_path $teacher \
  --save_steps 1000 \
  --dataset_name $dataset \
  --fp16 \
  --per_device_train_batch_size $real_bs \
  --gradient_accumulation_steps 4 \
  --train_n_passages 2 \
  --learning_rate $lr \
  --q_max_len 16 \
  --p_max_len 128 \
  --num_train_epochs $epoch \
  --logging_steps 500 \
  --overwrite_output_dir \
  --dataloader_num_workers 4 \