wantsleep

/

OLMoE_1B_7B_Eagle3

like 2

I used A6000 GPU x 2 40GB for training this model. It took one day (maybe 22hours?)

Got it! Thank you.

Hello,
I am currently training a model and would like to inquire about the expected accuracy and loss values during training.
After 2 epochs, my test accuracy and loss (pLoss) across different positions are as follows:

Test Epoch [2/40], position 0, Acc: 0.64
Test Epoch [2/40], position 1, Acc: 0.60
Test Epoch [2/40], position 2, Acc: 0.58
Test Epoch [2/40], position 3, Acc: 0.57
Test Epoch [2/40], position 4, Acc: 0.56
Test Epoch [2/40], position 5, Acc: 0.54
Test Epoch [2/40], position 6, Acc: 0.53
Test Epoch [2/40], position 0, pLoss: 1.57
Test Epoch [2/40], position 1, pLoss: 1.72
Test Epoch [2/40], position 2, pLoss: 1.81
Test Epoch [2/40], position 3, pLoss: 1.88
Test Epoch [2/40], position 4, pLoss: 1.94
Test Epoch [2/40], position 5, pLoss: 2.00
Test Epoch [2/40], position 6, pLoss: 2.08

For reference, my training configuration uses a maxlen of 4096, resulting in a total dataset size of 57,452 samples.
Could you please share what the typical or expected accuracy was in your training runs? I would like to know if my current results are reasonable or if they might indicate a potential issue.
Thank you for your time and help.