How to set save_steps? I set a 15,000, and mine has more than 34,000 steps. I just trained 1.28epoch at first sight, and directly generated 57GB checkpoints for me to dry up the hard disk.

[yuanllong]

Reminder

• I have read the README and searched the existing issues.

System Info

output

output_dir: /root/autodl-tmp
logging_steps: 10
save_steps: 15000
plot_loss: true
overwrite_output_dir: true

How to set save_steps? I set a 15,000, and mine has more than 34,000 steps. I just trained 1.28epoch at first sight, and directly generated 57GB checkpoints for me to dry up the hard disk.

Reproduction

output

output_dir: /root/autodl-tmp
logging_steps: 10
save_steps: 15000
plot_loss: true
overwrite_output_dir: true

How to set save_steps? I set a 15,000, and mine has more than 34,000 steps. I just trained 1.28epoch at first sight, and directly generated 57GB checkpoints for me to dry up the hard disk.

Others

output

output_dir: /root/autodl-tmp
logging_steps: 10
save_steps: 15000
plot_loss: true
overwrite_output_dir: true

How to set save_steps? I set a 15,000, and mine has more than 34,000 steps. I just trained 1.28epoch at first sight, and directly generated 57GB checkpoints for me to dry up the hard disk.

[hiyouga]

You can increase save_steps to 999999 or set save_strategy: no

[yuanllong]

您可以将 save_steps 增加到 999999 或设置 save_strategy：否

Oh, ok, thank you for your reply, so what is the behavior of saving in the case of save_steps=999999 or no? Only the best model will be saved? Or just save the last model at the end of the training?

[hiyouga]

save at the end of training