Train loss, dev loss, test loss

[YiWang31219]

Hi,

Thank you for sharing the PTEC code and dataset — I’ve been able to successfully reproduce your results using the original hatespeech dataset.

One observation I’d like to ask about is the training vs validation/test loss.
In my runs, the losses evolve approximately as follows:

• Training loss: starts at ~151 and eventually converges to ~21
• Dev loss: stabilizes around ~9
• Test loss: stabilizes around ~5

Is this behavior something you also observed?

If so, do you have any thoughts on what might contribute to this — e.g., soft prompt behavior, optimization setup, or something data-related?

Thanks in advance!

Best regards,
Yi