RWDE

RWDE is a webpage information extraction dataset built from English Common Crawl pages. Pages are organized by website and page-structure group, with extraction queries, JSON schemas, reference XPath locators, model-generated supervision, and field-level ground truth.

The dataset contains two parts:

• SFT: data used for supervised fine-tuning, including node-identification artifacts and generated scraper programs.
• RFT: data used for reinforcement fine-tuning, including answer-item and support-item annotations.

The current release excludes queries that lack extracted HTML outputs in SFT or item-level ground truth in RFT.

Statistics

Split	Sites	Groups	Queries	Raw HTML pages	Query-pages	Extracted HTML files	Scraper samples
SFT	151	367	548	26,469	38,919	30,359	46,585
RFT	33	95	95	3,826	3,826	2,777	—
Total	184	462	643	30,295	42,745	33,136	46,585

A query-page is one page evaluated under one query. A raw HTML page appearing in multiple queries is counted once for each query. RFT additionally contains 3,396 item-level ground-truth JSON files.

Download and extraction

The Hugging Face release contains three archives plus this README, a machine-readable manifest, and SHA-256 checksums:

SFT.tar.gz
RFT.tar.gz
groundtruth.tar.gz
manifest.json
SHA256SUMS

Each archive contains its corresponding top-level directory. To reconstruct the complete dataset:

sha256sum -c SHA256SUMS
tar -xzf SFT.tar.gz
tar -xzf RFT.tar.gz
tar -xzf groundtruth.tar.gz

The command sha256sum -c SHA256SUMS must be run from the directory containing the downloaded release files, before moving or extracting the archives.

Directory structure

RWDE/
├── README.md
├── SFT/
│   └── <website>/<group_id>/
│       ├── query.jsonl
│       ├── raw_html/
│       │   └── XXXX.html
│       ├── simplified_html/
│       │   └── XXXX_simplified.html
│       ├── mapped_html/
│       │   └── XXXX_mapped.html
│       └── extracted_html/
│           └── <query_id>/
│               ├── XXXX_extracted.html
│               ├── XXXX_ids.json
│               └── scraper.jsonl
├── RFT/
│   └── <website>/<group_id>/
│       ├── query.jsonl
│       ├── raw_html/
│       ├── simplified_html/
│       ├── mapped_html/
│       └── extracted_html/
│           ├── <query_id>/
│           │   ├── XXXX_extracted.html
│           │   └── XXXX_ids.json
│           └── <query_id>_gt/
│               └── XXXX.json
└── groundtruth/
    ├── SFT/<website>/<group_id>/<query_id>/<field_name>.txt
    └── RFT/<website>/<group_id>/<query_id>/<field_name>.txt

scraper.jsonl is present only for queries with generated scraper samples. In SFT, xxxtickling.com/group000/q01 and xxxtickling.com/group000/q02 do not have this file; the other 546 SFT queries do.

Query definitions

Each group has a query.jsonl file with one JSON object per query. Queries are selected from the passing audit results.

{
  "query_id": "q01",
  "query": "Extraction query in natural language",
  "schema": {
    "type": "object",
    "properties": {}
  },
  "reference_xpath": {
    "field_name": {
      "selectors": [
        {
          "xpath": "//main//h1",
          "value_mode": "text",
          "attribute": null
        }
      ],
      "combine": "direct",
      "separator": null
    }
  }
}

• schema: JSON Schema for the expected extraction result.
• reference_xpath: field name to locator configuration. A field can have multiple selectors and a composition rule.
• value_mode: text extracts visible node text; attribute extracts the named attribute.

HTML and node-identification artifacts

• raw_html/XXXX.html: original page HTML.
• simplified_html/XXXX_simplified.html: simplified HTML with item identifiers used during preprocessing.
• mapped_html/XXXX_mapped.html: mapped preprocessing representation.
• extracted_html/<query_id>/XXXX_extracted.html: query-specific HTML retained after node selection.
• extracted_html/<query_id>/XXXX_ids.json: model output for node identification.

Each XXXX_ids.json contains exactly:

{
  "reasoning": "Model reasoning",
  "output": "Raw model output",
  "useful_item_ids": ["17", "19", "24"]
}

Scraper samples

For SFT, scraper.jsonl contains one generated scraper sample per line with exactly these fields:

{
  "uid": "website/group000/q01/0000-0013-0032",
  "domain": "website",
  "group": "group000",
  "query_id": "q01",
  "seed_page_ids": ["0000", "0013", "0032"],
  "reasoning": "Model reasoning",
  "output": "Generated scraper source code"
}

Rows are preserved from the source artifacts. The same uid can occur multiple times with different reasoning or output, so uid is not guaranteed to be unique. Review generated code before executing it.

RFT item-level ground truth

In RFT, extracted_html/<query_id>_gt/XXXX.json aligns fields with annotated HTML items:

{
  "answer_item_ids": {
    "field_name": [17, 19]
  },
  "support_item_ids": [23, 27]
}

• answer_item_ids: item IDs containing direct answers for each field.
• support_item_ids: nearby or page-level items providing labels, structure, or disambiguating context.

Field-level ground truth

Ground truth is split into groundtruth/SFT and groundtruth/RFT, then organized by website, group, query, and field. SFT values are materialized with the reference XPath locators. RFT values come from the extracted standard answers.

Each <field_name>.txt uses an SWDE/LiveWeb-IE-style text format. The first two lines are headers; data rows begin on line 3:

SFT 24karotgoldcasino.com group000 q02 article_title
173 173 173
0000 "http://24karotgoldcasino.com/2018/06/15/washington-casinos-2/" 1 Washington Casinos
0001 "http://24karotgoldcasino.com/2010/04/22/wetten-a-atlantic-city-2/" 1 Wetten a Atlantic City

Each data row has this form:

<page_id> "<page_url>" <value_count> <value_1>\t<value_2>\t...

• Scalar fields normally have value_count 1.
• Array fields contain multiple values separated by tab characters.
• Missing scalar values and empty arrays have value_count 0 and no payload.
• Values are stored as text; use query.jsonl for their expected JSON types.

The two header lines contain:

1. <split> <website> <group_id> <query_id> <field_name>
2. <page_count> <total_value_count> <total_value_count>

License

The dataset compilation, query definitions, ground-truth annotations, and model-generated artifacts are released under the Creative Commons Attribution 4.0 International License.

You may copy, use, modify, and redistribute these materials, including for commercial purposes, provided that you give appropriate credit, provide a link to the license, and indicate whether you made changes. When using RWDE, please cite the associated paper and this Hugging Face repository.

Raw webpages and other third-party content remain subject to the rights and terms of their original owners. The CC BY 4.0 license applies only to materials for which the RWDE authors can grant rights and does not grant additional rights to third-party webpage content.

Page ID mapping

The four-digit page ID links every representation of the same page. For example, page 0000 can map to:

raw_html/0000.html
simplified_html/0000_simplified.html
mapped_html/0000_mapped.html
extracted_html/<query_id>/0000_extracted.html
extracted_html/<query_id>/0000_ids.json
extracted_html/<query_id>_gt/0000.json

Query-specific files exist only when the corresponding page was retained or annotated for that query. In every field-level ground-truth file, the same page ID is followed by the source page URL.