RefinedGAW

A Refined German Academic Web Dataset

Paper | Rehydration tool

RefinedGAW is a 314B token dataset based on 26 crawls from the German Academic Web. This repository contains a metadata index with rich annotations for 121.9 million unique HTML and PDF texts, incl. document-level content annotations, language labels, token counts, source URLs, and crawl provenance.

• 121.9M unique text documents (110M from HTML, 11M from PDFs)
• 314.5B tokens, with 235B in documents labeled as academic research or scientific
• Specialized content at substantial scale. The collection is especially rich in academic, scientific, educational, math-heavy, and expert-oriented material.
• Complementary coverage. Even under recall-oriented near-duplicate matching, 88% of German RefinedGAW documents have no candidate match in FineWeb2-DE, FinePDFs-DE, or German Commons, providing evidence that RefinedGAW adds substantial coverage beyond these established resources.
• Selection-ready metadata. Every record includes document length, two tokenizer-specific token counts, language predictions, a one-sentence description, and 17 structured content-property annotations.
• HTML and PDF coverage. Separate configurations expose web pages and PDF documents processed through format-specific extraction pipelines.
• Direct access to public sources. Source URLs and rehydration software support targeted retrieval and processing of documents that remain publicly available.

Public release: This repository contains document metadata and annotations, not the source text itself. The RefinedGAW metadata index and its derived annotations are released under CC BY 4.0. This license does not apply to the underlying source documents referenced by the index, which remain subject to their respective rights and licenses.

What does RefinedGAW contain?

Language composition

The GAW follows German academic institutions rather than a language boundary, so RefinedGAW is multilingual. German and English together account for 76% of HTML documents and 65% of HTML tokens. The corresponding PDF shares are 77% and 79%. Language labels are the top predictions from GlotLID v3 and are not human-verified.

Long documents

RefinedGAW contains a substantial amount of long documents. Documents containing at least 128k tokens contribute 57.7 billion HTML tokens and 20.7 billion PDF tokens. At a threshold of one million tokens, 19.6 billion HTML tokens and 2.1 billion PDF tokens remain.

Dataset structure

RefinedGAW is organized into separate configurations for HTML pages and PDF documents.

Configuration	Unique texts	Indexed GPT-OSS tokens	Source format
html	111.3M	229.6B	Text extracted from HTML pages
pdf	10.6M	84.9B	Text extracted from PDF documents

Each row describes one extracted text after exact-text deduplication across all 26 crawls. When the same text occurs under multiple URLs or in multiple crawls, RefinedGAW retains one record and stores its occurrence and version history in metadata fields such as alt_urls, crawls, and url_versions.

Both configurations contain a single train split holding the complete index. The public files contain metadata rather than source text, so the token totals above describe the extracted texts represented by the index.

Schema

The two configurations share 36 fields, and the PDF configuration adds 14 fields describing the source file and extraction process.

Core, language, and provenance fields

Column	Type	Description
num_tokens_gpt_oss	uint32	Number of GPT-OSS tokens in the extracted source text.
num_tokens_gemma_3	uint32	Number of Gemma 3 tokens in the extracted source text.
num_chars	uint32	Number of characters in the extracted source text.
quality_score	uint8	Deterministic convenience ranking from 0 to 132 based on seven content annotations. It is not a calibrated quality estimate.
url	string	Primary source URL associated with the record.
full_doc_lid	string	Top document-level GlotLID v3 language prediction, or unknown when the text fails the language-identification precheck.
full_doc_lid_score	float32	GlotLID score for the top prediction.
second_doc_lid	string	Second-ranked GlotLID language prediction.
second_doc_lid_score	float32	GlotLID score for the second-ranked prediction.
lid_text_chars	uint32	Length of the normalized text supplied to language identification, capped at the first 40,000 source characters.
lid_alpha_bytes	uint32	Number of UTF-8 bytes occupied by alphabetic characters in the normalized language-identification text.
lid_alpha_ratio	float32	Ratio of alphabetic UTF-8 bytes to non-whitespace UTF-8 bytes in the normalized language-identification text.
crawl	string	Crawl identifier associated with the primary occurrence.
timestamp	string	Capture timestamp associated with the primary occurrence.
alt_urls	list<struct{url: string, timestamp: string}>	Other retained URLs and capture timestamps that yielded the same extracted text.
crawls	list<struct{crawl: string, count: uint32}>	Crawls in which the text occurred and the number of retained captures in each crawl.
url_versions	list<struct{id: string, timestamp: string}>	Identifiers and timestamps of textually different versions observed at the primary URL.
id	string	Content-derived identifier for the extracted text.

Document-level content annotations

These fields are predictions produced by propella-1-4b rather than human-verified labels.

Column	Type	Description
one_sentence_description	string	Model-generated one-sentence description of the document.
content_type	list<string>	Predicted document types or communicative modes, such as analytical or instructional content.
business_sector	list<string>	Predicted sectors represented by the document, including academic research.
technical_content	list<string>	Predicted technical-content categories, including scientific and mathematical material.
audience_level	string	Predicted level of background knowledge expected from the reader.
content_quality	string	Predicted overall content quality.
educational_value	string	Predicted educational value.
reasoning_indicators	string	Predicted strength of explanation, analysis, or other reasoning signals.
information_density	string	Predicted density of substantive information.
content_integrity	string	Predicted completeness or degradation of the extracted content.
content_ratio	string	Predicted balance between substantive content and navigation or boilerplate.
content_length	string	Model-assigned document-length category.
commercial_bias	string	Predicted degree of marketing or commercial bias.
time_sensitivity	string	Predicted dependence on time-sensitive or rapidly changing information.
content_safety	string	Predicted presence and severity of content-safety concerns.
pii_presence	string	Predicted presence and degree of personally identifiable information.
regional_relevance	list<string>	Predicted geographic regions relevant to the document.
country_relevance	list<string>	Predicted countries relevant to the document.

PDF-only fields

Column	Type	Description
num_pages	int32	Number of pages in the PDF.
pages_ok	int32	Number of pages successfully processed.
pages_failed	int32	Number of pages that failed during processing.
pdf_bytes	int64	Size of the PDF source representation in bytes.
quality_flags	list<string>	Document-level flags derived from PDF extraction diagnostics.
pdf_sha256	string	SHA-256 digest of the retained PDF byte representation.
pdf_byte_variants	uint64	Number of byte-distinct PDF files that produced the same extracted text.
conversion_status	string	Overall PDF conversion status.
page_status_counts	list<struct{status: string, count: int32}>	Counts of page-level processing outcomes.
models	list<struct{model: string, revision: string, page_count: int32}>	OCR or extraction models, revisions, and numbers of pages processed with each model.
source_archive	string	Source archive containing the processed PDF.
source_member	string	Member path of the PDF within the source archive.
processed_at	string	Timestamp of PDF processing.
text_sha256	string	Full SHA-256 digest of the extracted text.

Select and rehydrate a subset

RefinedGAW is designed so that users can select documents from the metadata before downloading any source content. The following example streams the HTML configuration, selects documents whose technical_content annotation contains math_heavy, and saves the first ten matches with only the fields required for rehydration.

from datasets import Dataset, load_dataset

html = load_dataset(
    "ellamind/RefinedGAW",
    "html",
    split="train",
    streaming=True,
    columns=[
        "id",
        "url",
        "alt_urls",
        "technical_content",
    ],
)

math_heavy = html.filter(
    lambda row: "math_heavy" in (row["technical_content"] or [])
)

rehydration_fields = ("id", "url", "alt_urls")
records = [
    {field: row[field] for field in rehydration_fields}
    for row in math_heavy.take(10)
]
Dataset.from_list(records).to_parquet("math-heavy-html.parquet")

The resulting Parquet file can be passed directly to the rehydration tool:

git clone https://github.com/maxidl/refinedgaw-rehydrate.git
cd refinedgaw-rehydrate
uv sync --frozen

uv run --frozen refinedgaw download ../math-heavy-html.parquet \
  --format html \
  --output downloads/

Downloaded files are stored under downloads/content/, while downloads/results.jsonl records the outcome for every requested document. The production HTML extraction pipeline uses Trafilatura and runs without a GPU:

cd extractors/html
uv sync --frozen

uv run --frozen refinedgaw-extract-html \
  ../../downloads/results.jsonl \
  --output ../../downloads/html-extraction/

Successfully extracted UTF-8 text is written under downloads/html-extraction/texts/. The tool retrieves only sources that remain publicly available, respects current robots.txt rules, and does not provide access to the original GAW archive.

Rehydration and source availability

The indexed documents were captured between 2012 and 2025, so present-day source availability varies substantially with age. In a stratified experiment conducted on August 3, 2026, an estimated 35.7% of the full HTML index and 40.4% of the full PDF index could still be downloaded from public source URLs. For documents last observed in the June 2025 crawl, the corresponding rates rose to 75.2% and 82.6%.

Most unsuccessful downloads were explained by missing resources, current robots.txt exclusions, and DNS failures. Present-day retrieval also does not guarantee exact reproduction of the archived extracted text because web pages change and PDF OCR can be nondeterministic. These measurements describe the state of the public web on the experiment date rather than guaranteed future availability.

Dataset construction

For details of the HTML and PDF extraction pipelines, exact-text deduplication, provenance aggregation, document annotation, validation, and corpus analysis, see the RefinedGAW paper.

Limitations and responsible use

RefinedGAW is a focused collection rather than a representative sample of German-language text or academic publishing. The GAW starts around 150 institutional seed domains, and its coverage reflects that selection, the institutions' link structures, and the crawl policy. The corpus is also concentrated by source: the ten largest HTML hosts contribute 36% of HTML tokens, so applications that require balanced coverage should sample or weight hosts explicitly.

Language labels and document-level content annotations are model predictions and may be incorrect. The quality_score is a convenience ranking derived from selected annotations, not a human judgment or a calibrated estimate of suitability. RefinedGAW has not been evaluated through language-model training, so annotation-based selection should be validated for the intended downstream use.

The public dataset contains metadata, not source text. Rehydration retrieves only currently available public sources, whose availability and contents can change over time, and does not guarantee reproduction of the archived extraction. Source documents may contain copyrighted material or personal information, including contact details and staff directories. The pii_presence field is an imperfect model prediction and should not be treated as a complete safeguard. Users are responsible for respecting applicable law, source licenses, website terms, privacy interests, and current robots.txt rules. The CC BY 4.0 license applies to the RefinedGAW metadata index and its derived annotations. It does not grant rights to the underlying source documents.

Privacy and record-removal requests

This repository contains metadata and model-generated annotations derived from public academic websites. It does not distribute archived HTML, PDF files, or extracted source text. Some records may nevertheless relate to identifiable people. If a record concerns you, or if you represent its source publisher or another relevant rights holder, contact the project privately at info [at] ellamind [dot] com with its RefinedGAW id or source URL to request review, correction, restriction, or removal. Requests will be assessed as appropriate. Please do not include unnecessary personal information.

Citation

RefinedGAW is derived from the German Academic Web. If you use this dataset, please cite both the RefinedGAW paper and the original GAW paper.

@misc{refinedgaw2026,
  title  = {RefinedGAW: A Refined German Academic Web Dataset},
  author = {Idahl, Maximilian and Astappiev, Oleh and Droste, Benedikt and Scholz, Robert and Pl{\"u}ster, Bj{\"o}rn and Harries, Jan Philipp and J{\"a}schke, Robert and Nejdl, Wolfgang},
  year   = {2026},
  eprint = {2610.34572},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url    = {https://arxiv.org/abs/2610.34572}
}

@inproceedings{paris2020gaw,
  title     = {How to Assess the Exhaustiveness of Longitudinal Web Archives: A Case Study of the {German Academic Web}},
  author    = {Paris, Michael and J{\"a}schke, Robert},
  booktitle = {Proceedings of the 31st ACM Conference on Hypertext and Social Media},
  pages     = {85--89},
  publisher = {Association for Computing Machinery},
  year      = {2020},
  doi       = {10.1145/3372923.3404836},
  url       = {https://doi.org/10.1145/3372923.3404836}
}