Recommended settings for CPU-only training with limited RAM

[RocketsX]

Hi! I’m evaluating Laya for a cryptocurrency market research project and would like to understand the safest configuration for training on a Windows machine with limited resources.

The dataset has about 24,520 training examples, with 60 timesteps of input features. The machine has limited free RAM, and training would run on CPU only.

Could you recommend:

• Whether --freeze-encoder is appropriate for this dataset and setup;
• Which batch size and other memory-related settings to start with;
• Whether training supports limiting CPU threads or memory use;
• How to monitor for model collapse or overfitting in this situation.

I’m currently doing research and validation only; no live trading is involved. I can provide sanitized logs or configuration details if useful.

Thanks!