florin-hf/kilt_corpus_wiki_dump2019

KILT Corpus in Chunks: Wikipedia Corpus Chunked for RAG

A preprocessed version of the KILT Wikipedia corpus where each article has been split into coherent, section-aware text chunks suitable for Retrieval Augmented Generation (RAG) pipelines.

Dataset Summary

This dataset provides a chunked view of the KILT Wikipedia snapshot (August 2019). Starting from full Wikipedia articles, each article is first segmented by its section structure, then each section is independently chunked using a sentence-aware splitter. This design ensures that chunks never cross section boundaries and that each chunk can be precisely located within its source article.

Processing Pipeline

1. Source: facebook/kilt_wikipedia — a Wikipedia snapshot from 01 August 2019 used as the knowledge base for the KILT benchmark.
2. Section splitting: Each article is divided into its constituent sections/subsections before chunking. No chunk ever spans two different sections.
3. Sentence-aware chunking: Each section is chunked using SentenceSplitter from LlamaIndex, producing non-overlapping chunks of up to 256 tokens. Sentence boundaries are respected, so chunks are always semantically coherent.
4. Chunk and position metadata are added to each record (see schema below).

---

Schema

Each record corresponds to one chunk and contains the following fields:

Field	Type	Description
chunk_id	string	Unique chunk identifier in the format doc-{doc_id}::s-{section_order}::chunk-{chunk_order}
doc_id	string	Document identifier combining Wikipedia ID and revision info
chunk_order	int	Position of the chunk within its section (0-indexed)
is_first	bool	Whether this is the first chunk in the section
is_last	bool	Whether this is the last chunk in the section
text	string	The chunk text content
title	string	Hierarchical section path (e.g. # Article\n## Section\n### Subsection)
section	list[str]	List of section/subsection names for the chunk's location
section_order	int	Index of the section within the article (0-indexed)
num_sections	int	Total number of sections in the article
wikipedia_id	string	Original Wikipedia page ID
wikipedia_title	string	Wikipedia article title
categories	string	Comma-separated Wikipedia categories
pageid	string	Wikipedia page ID
parentid	string	Wikipedia parent revision ID
revid	string	Wikipedia revision ID
timestamp	string	Revision timestamp (ISO 8601)
url	string	Direct URL to the Wikipedia revision
wikidata_id	string	Wikidata entity ID

Example Record

{
  'chunk_id': 'doc-39_4::s-1::chunk-0',
  'doc_id': '39_4',
  'chunk_order': 0,
  'is_first': True,
  'is_last': True,
  'text': "Albedo is not directly dependent on illumination because changing the amount "
          "of incoming light proportionally changes the amount of reflected light, except "
          "in circumstances where a change in illumination induces a change in the Earth's "
          "surface at that location (e.g. through albedo-temperature feedback). That said, "
          "albedo and illumination both vary by latitude. Albedo is highest near the poles "
          "and lowest in the subtropics, with a local maximum in the tropics.",
  'title': '# Albedo\n## Examples of terrestrial albedo effects\n### Illumination',
  'wikipedia_id': '39',
  'wikipedia_title': 'Albedo',
  'section': ['Examples of terrestrial albedo effects', 'Illumination'],
  'section_order': 4,
  'num_sections': 19,
  'url': 'https://en.wikipedia.org/w/index.php?title=Albedo&oldid=906500850',
  ...
}

---

Source & License

Built upon facebook/kilt_wikipedia. Please refer to the original dataset for licensing terms. The additional processing and metadata fields are released under the same terms.

This processed corpus was used in the papers Do RAG Systems Really Suffer From Positional Bias? and Redefining Retrieval Evaluation in the Era of LLMs.

@inproceedings{petroni-etal-2021-kilt,
    title = "{KILT}: a Benchmark for Knowledge Intensive Language Tasks",
    author = {Petroni, Fabio  and
      Piktus, Aleksandra  and
      Fan, Angela  and
      Lewis, Patrick  and
      Yazdani, Majid  and
      De Cao, Nicola  and
      Thorne, James  and
      Jernite, Yacine  and
      Karpukhin, Vladimir  and
      Maillard, Jean  and
      Plachouras, Vassilis  and
      Rockt{\"a}schel, Tim  and
      Riedel, Sebastian},
    editor = "Toutanova, Kristina  and
      Rumshisky, Anna  and
      Zettlemoyer, Luke  and
      Hakkani-Tur, Dilek  and
      Beltagy, Iz  and
      Bethard, Steven  and
      Cotterell, Ryan  and
      Chakraborty, Tanmoy  and
      Zhou, Yichao",
    booktitle = "Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
    month = jun,
    year = "2021",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.naacl-main.200/",
    doi = "10.18653/v1/2021.naacl-main.200",
    pages = "2523--2544",
    abstract = "Challenging problems such as open-domain question answering, fact checking, slot filling and entity linking require access to large, external knowledge sources. While some models do well on individual tasks, developing general models is difficult as each task might require computationally expensive indexing of custom knowledge sources, in addition to dedicated infrastructure. To catalyze research on models that condition on specific information in large textual resources, we present a benchmark for knowledge-intensive language tasks (KILT). All tasks in KILT are grounded in the same snapshot of Wikipedia, reducing engineering turnaround through the re-use of components, as well as accelerating research into task-agnostic memory architectures. We test both task-specific and general baselines, evaluating downstream performance in addition to the ability of the models to provide provenance. We find that a shared dense vector index coupled with a seq2seq model is a strong baseline, outperforming more tailor-made approaches for fact checking, open-domain question answering and dialogue, and yielding competitive results on entity linking and slot filling, by generating disambiguated text. KILT data and code are available at \url{https://github.com/facebookresearch/KILT}."
}


@misc{cuconasu2025ragsystemsreallysuffer,
      title={Do RAG Systems Really Suffer From Positional Bias?}, 
      author={Florin Cuconasu and Simone Filice and Guy Horowitz and Yoelle Maarek and Fabrizio Silvestri},
      year={2025},
      eprint={2505.15561},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2505.15561}, 
}


@misc{trappolini2025redefiningretrievalevaluationera,
      title={Redefining Retrieval Evaluation in the Era of LLMs}, 
      author={Giovanni Trappolini and Florin Cuconasu and Simone Filice and Yoelle Maarek and Fabrizio Silvestri},
      year={2025},
      eprint={2510.21440},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2510.21440}, 
}