About data split

[Snowilv]

Hi, thanks for this impressive work! How can I split the Amazon review dataset into training/validation/test sets? Sadly I didn't see related codes in the repo, is there any guideline?

[Re-bin]

Hi 👋

Thanks for you interests on RLMRec! The operation of splitting into training/validation/testing sets is performed on an interaction basis.

After completing filtering and k-core operations, let's assume we have 100,000 remaining user-item interaction records. If we require a ratio of 3:1:1 for training/validation/testing, we randomly sample 60,000 interactions to construct the training set (ensuring that each user and item have at least one interaction; otherwise, we resample). The remaining 40,000 interactions are randomly divided in half to create the validation and testing sets.

I hope the provided answers can address your questions :)

Best regards, Xubin

[Snowilv]

Thx! It works for me.

[Snowilv]

Sorry to bother again. The way to generate profile confuses me now.

So do we need to send the system prompt every time before we obtain a user/item information, or we do it every few times? Considering system prompt is really token consuming.

[Re-bin]

Hi 👋

Yeah, when generating profiles, we need to send the system prompt each time. This is because, when calling the API, each query is treated as a new conversation, requiring us to resend the system prompt. The system prompt will take around 400 tokens.

However, with the improved language model's understanding capabilities, I believe we can now try reducing unnecessary descriptions in the system prompt and see if the output still meets the desired requirements :)

Best regards, Xubin

[Snowilv]

Get it! Thanks for your thorough response. Happy New Year by the way!