File size: 1,829 Bytes
17ac2d5 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 |
Continued pretrained from the nb-roberta-base.
The domain specific pretraining is done on the 102GB (Scandinavian corpus)[https://huggingface.co/datasets/NbAiLab/scandinavian].
## Train for 180k steps for 128 sequences:
```bash
./run_mlm_flax_stream.py \
--output_dir="./" \
--model_type="roberta" \
--config_name="./" \
--tokenizer_name="./" \
--model_name_or_path="./" \
--dataset_name="NbAiLab/scandinavian" \
--max_seq_length="128" \
--weight_decay="0.01" \
--per_device_train_batch_size="128" \
--per_device_eval_batch_size="128" \
--learning_rate="6e-5" \
--warmup_steps="5000" \
--overwrite_output_dir \
--cache_dir /mnt/disks/flaxdisk/cache/ \
--num_train_steps="180000" \
--adam_beta1="0.9" \
--adam_beta2="0.98" \
--logging_steps="10000" \
--save_steps="10000" \
--eval_steps="10000" \
--preprocessing_num_workers 96 \
--auth_token True \
--adafactor \
--push_to_hub
```
## Train for 20k steps for 512 sequences:
```bash
./run_mlm_flax_stream.py \
--output_dir="./" \
--model_type="roberta" \
--config_name="./" \
--tokenizer_name="./" \
--model_name_or_path="./" \
--dataset_name="NbAiLab/scandinavian" \
--max_seq_length="512" \
--weight_decay="0.01" \
--per_device_train_batch_size="48" \
--per_device_eval_batch_size="48" \
--learning_rate="3e-5" \
--warmup_steps="5000" \
--overwrite_output_dir \
--cache_dir /mnt/disks/flaxdisk/cache/ \
--num_train_steps="20000" \
--adam_beta1="0.9" \
--adam_beta2="0.98" \
--logging_steps="20000" \
--save_steps="10000" \
--eval_steps="10000" \
--preprocessing_num_workers 96 \
--auth_token True \
--adafactor \
--push_to_hub
```
Approximate additional training time: 1 week.
|