siavashsaki/wdc-pave-ave

WDC-PAVE: Attribute-Value Extraction Benchmark

A cleaned, canonicalized, and pre-split version of the WDC Product Attribute-Value Extraction (PAVE) dataset, prepared for structured information extraction experiments with LLMs.

Task

Given a product title and description, extract attribute-value pairs into a JSON object with a fixed schema per product category.

• Input: product title + product description (free text)
• Output: JSON object with category-specific attributes; missing values are null

This is a structured extraction task, not open-ended generation. Each category has a fixed set of expected attributes.

Dataset Details

Source	WDC-PAVE (normalized variant)
Records	1,420
Categories	5
Unique attributes	24 (3-11 per category)
Split	70% train / 10% val / 20% test, stratified by category
Random seed	42

Category distribution

Category	Train	Val	Test	Total	Attributes
Computers And Accessories	305	44	87	436	11
Home And Garden	250	35	71	356	8
Office Products	207	30	60	297	10
Jewelry	175	25	50	250	3
Grocery And Gourmet Food	57	8	16	81	5
Total	994	142	284	1,420

Per-category schemas

Computers And Accessories (11 attributes): Generation, Part Number, Product Type, Cache, Processor Type, Processor Core, Interface, Manufacturer, Capacity, Ports, Rotational Speed

Home And Garden (8 attributes): Product Type, Color, Length, Width, Height, Depth, Manufacturer Stock Number, Retail UPC

Office Products (10 attributes): Product Type, Color(s), Pack Quantity, Length, Width, Height, Depth, Paper Weight, Manufacturer Stock Number, Retail UPC

Jewelry (3 attributes): Product Type, Brand, Model Number

Grocery And Gourmet Food (5 attributes): Product Type, Brand, Pack Quantity, Retail UPC, Size/Weight

Record format

Each JSONL record has the following fields:

{
  "id": 8068358,
  "category": "Home And Garden",
  "input_title": "Pneumatic Lift Lab Stools w/Back ...",
  "input_description": "Pneumatic lift adjusts to accommodate ...",
  "gold_json": {
    "Product Type": "Furniture, Storage, Racks and Fixtures",
    "Color": "Black",
    "Length": null,
    "Width": null,
    "Height": "99.1",
    "Depth": null,
    "Manufacturer Stock Number": "SAF3430BL",
    "Retail UPC": null
  }
}

Value conventions

• Single value -> string: "Manufacturer": "Dell"
• Multiple values -> sorted list: "Manufacturer": ["Hewlett-Packard", "Hewlett-Packard Enterprise"]
• Missing / not applicable -> null
• Multi-value attributes occur in 3.9% of attribute instances

Data preparation

The following cleaning steps were applied to the raw WDC-PAVE normalized variant:

1. Title cleaning: Stripped literal "Null" scraping artifacts from 368 product titles (26% of records)
2. Gold canonicalization: Parsed target_scores structure into flat JSON per category schema; converted n/a to null; sorted multi-value lists alphabetically
3. Normalization: Trimmed whitespace, collapsed repeated spaces. Casing is preserved (use case-insensitive comparison at evaluation time)
4. Schema enforcement: Each record's gold_json contains all attributes for its category, with null for absent values

Suggested evaluation metrics

1. Valid JSON rate -- can the output be parsed?
2. Schema adherence -- correct keys, no extras, valid types?
3. Field-level precision / recall / F1 -- case-insensitive, set-level matching for multi-value attributes
4. Object exact match -- full JSON matches gold?
5. Null handling -- hallucination rate, miss rate, null accuracy
6. Latency and cost per 1,000 examples

Citation

If you use this dataset, please cite the original WDC-PAVE paper:

@inproceedings{primpeli2023wdcpave,
  title={Product Attribute Value Extraction using Large Language Models},
  author={Primpeli, Anna and Bizer, Christian},
  year={2023}
}

License

This dataset inherits the CC BY 4.0 license from the original WDC-PAVE dataset.