Downloads · 30 days
9
2% of all-time downloads
crystina-z/marginmse.teacher-monoXLMR.pft-msmarco.epoch-5.32x2.lr.7e-6.pft-msmarco
marginmse.teacher-monoXLMR.pft-msmarco.epoch-5.32x2.lr.7e-6.pft-msmarco is a feature extraction model from crystina-z. Use it when you need embeddings to search or compare text. It is set up for transformers.
Downloads · 30 days
9
2% of all-time downloads
All-time downloads
406
Public
Repo size
1.4 GB
Likes
0
Public
Click a slice to open those files.
.bin711 MB · 99%
From the Hugging Face model README
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 \