README.md

---
dataset_info:
  features:
  - name: id
    dtype: int64
  - name: text
    dtype: string
  - name: title
    dtype: string
  splits:
  - name: train
    num_bytes: 13638576512
    num_examples: 20970784
  download_size: 7557029888
  dataset_size: 13638576512
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
task_categories:
- question-answering
language:
- en
size_categories:
- 10M<n<100M
---
# Wikipedia Dump without Duplicates

## Dataset Summary
This is a cleaned and de-duplicated version of the English Wikipedia dump dated December 20, 2018. Originally sourced from the [DPR repository](https://github.com/facebookresearch/DPR), it has been processed to remove duplicates, resulting in a final count of **20,970,784** passages, each consisting of 100 words.
The original corpus is available for download via this [link](https://dl.fbaipublicfiles.com/dpr/wikipedia_split/psgs_w100.tsv.gz).

The corpus is used in the research paper [A Tale of Trust and Accuracy: Base vs. Instruct LLMs in RAG Systems](https://arxiv.org/abs/2406.14972), supporting experiments comparing base and instruct Large Language Models within Retrieval-Augmented Generation systems.