Datasets:

iam-tsr
/

ragmix

language:
  - en
license: apache-2.0
pretty_name: RAGmix
task_categories:
  - question-answering
  - text-retrieval
task_ids:
  - extractive-qa
  - closed-domain-qa
tags:
  - rag
  - retrieval-augmented-generation
  - evaluation
  - benchmark
  - multi-domain
  - heterogeneous
  - document-qa
  - apache-2.0
  - digitalcorpora
configs:
  - config_name: Ragmix
    data_files:
      - split: test
        path: data/test-00000-of-00001-43be72dace544f4fba6c2a484a3a0032.parquet
size_categories:
  - n<1K
dataset_info:
  features:
    - name: category
      dtype: string
    - name: document
      dtype: string
    - name: question
      dtype: string
    - name: answer
      dtype: string
  splits:
    - name: test
      num_examples: 100

RAGmix

RAGmix is a heterogeneous, multi-domain evaluation dataset for Retrieval-Augmented Generation (RAG) systems.

It mixes real-world document styles—policies, meeting minutes, clinical and scientific text, financial disclosures, job postings, and more—so models can be tested outside a single vertical. Each example pairs a full source document with one grounded question and a reference answer.

Source PDFs were obtained from Digital Corpora and converted to markdown for this release. This is a test-only split (100 examples).

How the data was built

Documents — Heterogeneous PDFs were downloaded from Digital Corpora, a public repository of digital corpora for forensics education and research, and converted to markdown (OCR/layout-aware parsing with image analysis where needed).
Questions — One evaluation question per document, targeting specific, document-grounded facts (dates, amounts, requirements, findings, procedures).
Answers — Short reference answers written from the same document only (no external knowledge required for the gold answer).
Categories — Manual topic labels for multi-domain analysis.

from datasets import load_dataset
rag_dataset = load_dataset("iam-tsr/ragmix")

Limitations

Small scale (100 examples); statistical significance is limited.
One question per document; does not cover multi-hop or multi-document reasoning.
Documents vary widely in length and quality (OCR/markdown artifacts may remain).
Categories are coarse; some documents sit near domain boundaries.
English only.
Reference answers are human-written summaries of document facts; alternative phrasings may also be correct.

Ethical considerations

Document texts originate from third-party PDFs hosted via Digital Corpora; treat redistribution of full text carefully under applicable rights and Digital Corpora’s terms.
Some documents touch sensitive topics (health, religion, legal enforcement). Prefer evaluation / research use.
Do not treat answers as professional medical, legal, or financial advice.

Citation

@misc{ragmix,
  title        = {RAGmix: A Heterogeneous Multi-Domain Dataset for RAG Evaluation},
  author       = {Tushar Soni},
  year         = {2026},
  howpublished = {Hugging Face Datasets},
  note         = {Test split, 100 document-grounded QA examples. Source PDFs from Digital Corpora (https://digitalcorpora.org/).}
}

@misc{digitalcorpora,
  title        = {Digital Corpora},
  author       = {Garfinkel, Simson L. and others},
  howpublished = {\url{https://digitalcorpora.org/}},
  note         = {Public digital corpora for education and research}
}

License

This public extract is made available under Apache license 2.0. Users should also abide to the Digital Corpora.

Changelog

v0.1 — Initial test release: 100 examples, 15 categories.