Keyword arguments {'add_prefix_space': False} not recognized.

[rabeehkarimimahabadi]

Hi
I am getting this message when running finetune_trainer.py on TPUs any idea how to remove it? thanks

[LysandreJik]

Hi @rabeehkarimimahabadi. Could you please respect the issue template when opening issues? It is necessary for us to help you, otherwise we'll have to ask again and again for you to provide more information.

Please complete the issue template with the information related to your environment, to the command you launched, and the full error with the stack trace. Thanks.

[rabeehkarimimahabadi]

Hi I am sorry here is the full command I use , it is running finetune_trainer with T5 model. There is not an error, only it gives this warning. thanks python xla_spawn.py --num_cores 8 finetune_trainer.py --learning_rate=3e-5 --n_val 1000 --model_name_or_path t5-small --data_dir data/wmt_en_de/ --output_dir /home/rabeeh/temp/ --overwrite_output_dir --tpu_num_cores=8 --max_source_length=64 --max_target_length=64 --per_device_train_batch_size=32 --per_device_eval_batch_size=32 --label_smoothing=0.1 --task="translation" --logging_steps=200 --eval_steps=500 --num_train_epochs=6 --save_steps=500 --max_source_length=128 --max_target_length=128 --val_max_target_length=128 --test_max_target_length=128

[LysandreJik]

Since you're using the T5 model I believe you can safely ignore that warning, which is useful only for the byte-level BPE tokenizers (think GPT-2, RoBERTa).

Maybe @patil-suraj or @sshleifer know why such a warning is thrown when using T5 with the Seq2SeqTrainer

[sshleifer]

Yeah it's happening in seq2seq dataset, you can safely ignore. Fixed in #8329

[rabeehkarimimahabadi]

Hi, thank you. Is this applied to the repo? thanks

…

On Thu, Nov 5, 2020 at 4:15 PM Sam Shleifer ***@***.***> wrote: Yeah it's happening in seq2seq dataset, you can safely ignore. Fixed in #8329 <#8329> — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#8322 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ARPXHH2WLTMU6NJ7B2K2Y33SOK6SZANCNFSM4TLKSTWQ> .

[LysandreJik]

Yes, the fix has just been merged.

[rabeehkarimimahabadi]

Hi there, this is not fixed for me. and still I am getting it with running T5 using finetune_trainer.

[LysandreJik]

Have you pulled from master to get the updated version?

In any case, as I've said before, this is a warning and not an error. You can safely ignore it.