Available Datasets

AmazonDataset

The Amazon Dataset is a processed implementation of the Amazon Dataset provided by the author's of the P5 Paper. There are three available splits (beauty, sports, toys) with the following statistics (image taken from the mentioned P5 paper):

For each user, the following information is available:

• The sequence of bought items
• The rating assigned to each item

For each item, the following information is available:

• Description
• Categories
• Title
• Price
• Image URL
• Brand

AmazonDataset parameters

AmazonDataset:

  # The Amazon dataset split to use. Can be 'beauty', 'sports', 'toys'
  #
  # Required
  dataset_name: toys

  # If set to true, users and items ids will have "user_" and "item_" prefix in the dataset
  # E.g.:
  # add_prefix_items_users: false -> users = [1, 2, 3, ...]; items = [1, 2, 3, ...]
  # add_prefix_items_users: true -> users = [user_1, user_2, user_3, ...]; items = [item_1, item-2, item_3, ...]
  #
  # Optional, Default: true
  add_prefix_items_users: true

  # If set to true item ids will start from 1001 rather than 1  # (1)
  # E.g.:
  # items_start_from_1001: false -> items = [1, 2, 3, ...]
  # items_start_from_1001: true -> items = [1001, 1002, 1003, ...]
  items_start_from_1001: false

1. This is to fully exploit the LLM model tokenization: with sequential indexing, items with similar id should have more importance, thus by starting item ids from 1001 rather than 1 the sentencepiece tokenizer will tokenize with same subtokens items with similar ids! For more details check the following paper