Discussion on train_loss Behavior during Quiet-STaR Project Replication
Hello,
I’m currently attempting to replicate Quiet-STaR, and I've encountered a pattern in the training loss (train_loss) metric that I'm hoping to better understand with your guidance.
In my training, I've observed that the train_loss initially spikes, then decreases sharply, and finally oscillates around a certain value. This behavior seems somewhat unusual, as I would typically expect the train_loss to decrease gradually and stabilize over time. I’ve reviewed the data and parameters in line with the project recommendations, but the cause of this pattern still eludes me.
Could you provide insights into whether this is expected behavior, or suggest specific parameters or implementation details that might help address it?
Thank you very much for your time and any help you can provide!