dwb2023/ragas-golden-dataset-documents

Dataset Card for RAGAS Golden Dataset Documents

A small, mixed‐format corpus to compare PDF, API, and web‐based document loader output from the LangChain ecosystem.

The code to run the Prefect prefect_docloader_pipeline.py pipeline is available in the RAGAS Golden Dataset Pipeline repository.

While several enhancements are planned for future iterations, the hands-on insights gained from this grassroots exploration of document loader behaviors proved too valuable -- things that become less obvious with additional complexity.

The "aha" moment
I was shocked to discover how dramatically the choice of extraction method affects both content quality and metadata richness. These differences directly impact retrieval accuracy in ways I never anticipated until seeing the evidence side-by-side.

Dataset Description

This dataset contains 56 examples drawn from three ArXiv preprints (IDs: 2505.10468, 2505.06913, 2505.06817) in May 2025, each loaded via:

• PDF pages (50 records — one per page)
• ArXiv metadata summaries (3 records)
• Web HTML dumps (3 records)

Each item includes:

• page_content — the extracted text (page or summary)
• metadata_json — a JSON‐string of only non‐empty metadata fields

Why this dataset?
It lets you deep-dive on how different loader backends (PDF vs. arXiv API vs. web scrape) affect text extraction and metadata completeness in RAG ingestion pipelines.

Provenance

• Source: arXiv.org preprints (May 2025)

• Loaders:

  • PyPDFDirectoryLoader for PDF
  • ArxivLoader.get_summaries_as_docs() for API
  • WebBaseLoader for HTML

• Processing:

  1. Tag each document with its loader type (pdf / arxiv / webbase).
  2. Prune out empty metadata fields.
  3. Serialize remaining metadata to a single JSON string.
  4. Bundle into a Hugging Face Dataset via datasets.Dataset.from_list().

Dataset Structure

Field	Type	Description
page_content	string	Full text extracted from the page/summary.
metadata_json	string	JSON‐encoded flat dict of non‐empty metadata.

All examples live in the train split (56 total).

Intended Uses

• Benchmarking: Compare loader accuracy and completeness.
• Pipeline Prototyping: Rapidly test ingestion logic in RAG systems.
• Filtering Experiments: Use loader‐specific metadata to explore retrieval strategies.

Out-of-Scope

• Large-scale LLM training (too small).
• Tasks requiring human annotations or sensitive data.

Limitations & Risks

• Domain bias: Only three computer-science preprints, same month.
• Scale: Toy-scale corpus... unsuitable for production training.

Citation

@misc{branson2025ragas_gold_docs,
  title        = {RAGAS Golden Dataset Documents},
  author       = {Branson, Don},
  year         = {2025},
  howpublished = {\\url{https://huggingface.co/datasets/dwb2023/ragas-golden-dataset-documents}},
}

Maintainers & Contact

• Maintainer: Don Branson dwb2023
• Repository: https://github.com/donbr/ragas-golden-dataset-pipeline
• Hub Page: https://huggingface.co/datasets/dwb2023/ragas-golden-dataset-documents