Size of the dev and test sets

Old way of splitting data

• 70% Training
• 30% Test

or

• 60% training
• 20% development for iternations
• 20% test

These splits were reasonable when dataset is relatively small (100-10,000 samples)

If you have 1,000,000 samples, then you could use

• 98 % Training
• 2 % dev
• 2 % test

Size of test set

• Set your test set to be big enough to give high confidence in teh overall performance of your system

Depends on applications, you may not need test set

More discussion here