README.md

metadata

dataset_info:
  - config_name: pairs
    features:
      - name: query
        dtype: string
      - name: document
        dtype: string
      - name: relevance
        dtype: float64
      - name: source
        dtype: string
    splits:
      - name: train
        num_bytes: 2565164850
        num_examples: 5571429
      - name: test
        num_bytes: 730814746
        num_examples: 1462128
    download_size: 1234904598
    dataset_size: 3295979596
  - config_name: triplets
    features:
      - name: anchor
        dtype: string
      - name: positive
        dtype: string
      - name: negative
        dtype: string
      - name: margin
        dtype: float64
      - name: source
        dtype: string
      - name: metadata
        dtype: string
    splits:
      - name: train
        num_bytes: 26826437400
        num_examples: 28941558
      - name: test
        num_bytes: 26826437400
        num_examples: 6722877
    download_size: 583323916
    dataset_size: 2399240463
configs:
  - config_name: pairs
    data_files:
      - split: train
        path: pairs/train-*
      - split: test
        path: pairs/test-*
  - config_name: triplets
    data_files:
      - split: train
        path: triplets/train*
      - split: test
        path: triplets/test*
license: apache-2.0

This product search dataset compiles multiple open source product search datasets, that can be used for representation learning tasks.

Sources

Dataset	Repo ID	Source
Google	Marqo/marqo-GS-10M	Google Shopping
Amazon	tasksource/esci	Amazon ESCI
Wayfair	napsternxg/wands	Wayfair
Home Depot	bstds/home_depot	Home Depot
Crowdflower	napsternxg/kaggle_crowdflower_ecommerce_search_relevance	Crowdflower

Schema

Document

To standardize attributes across different sources and their availability, we use a template that can be applied based on available product information.

if kwargs.get("title"):
    template = f"""**product title**: {kwargs.get('title')}\n"""
else:
    template = """"""
if kwargs.get("category"):
    template += f"""**product category**: {kwargs.get('category').replace(" / ", " > ")}\n"""
if kwargs.get("attributes"):
    template += """**product attributes**:\n"""
    for k, v in kwargs.get("attributes").items():
        template += f""" - **{k}**: {v}\n"""

if kwargs.get("description"):
    template += f"""**product description**: {kwargs.get('description')}"""

The dataset has two splits:

• Pairs
• Triplets

Pairs

Query: The user query. Document: The product that was retrieved by the system. Relevance: The relevance of the <query, document> pair.

Each individual source will have their logic for sampling queries, documents, and relevance assessments. Most of the sources and manually graded by a group of annotators, except for Marqo/marqo-GS-10M which is the top 100 products retrieved from the system. I recommend reading the individual sources for a deeper understanding of their methodology.

This format undergoes no filtering, and all <query, document, relevance> scores are maintained from the original source. These can be directly used for training sentence similarity tasks that uses <sentence 1, sentence 2, score>. The scores should generally follow the range of 0-3, normalized across sources, but are not fully calibrated for the individual distributions.

Triplets

Train

Dataset	Queries	Documents	Pairs
Google	77,288	2,202,907	3,926,764
Amazon	99,408	985,476	1,420,372
Wayfair	477	38,854	140,068
Home Depot	11,795	54,360	74,067
Crowdflower	261	9,912	10,158

Test

Dataset	Queries	Documents	Pairs
Google	19,564	748,386	981,204
Amazon	30,947	364,004	434,234
Wayfair	477	25,317	46,690