|
--- |
|
license: cc-by-sa-3.0 |
|
--- |
|
# Dataset Card for Dataset Name |
|
|
|
This is a Wikipedia dataset correct to "31-12-2020". |
|
|
|
|
|
## Dataset Details |
|
|
|
### Dataset Description |
|
|
|
WikiMedia routinely publishes dumps of Wikipedia, each containing the revision history of articles. We first defined the relevant revision before extracting the article information. Specifically, we select the most recent revision as of December 31st for each year. Consequently, some revisions in our datasets date back several years from the target date since these pages haven't been edited. While this inclusion of older revisions might initially appear problematic, it is important to note that these are the existing versions of Wikipedia pages as of the cutoff date. The content of these pages was considered current enough at that time. This approach ensures that our training datasets reflect the most up-to-date information available on Wikipedia at each year's end, providing a realistic snapshot of knowledge for that specific point in time. |
|
|
|
Once each revision has been identified we clean the page using the code from \textit{wiki-dump-reader} \footnote{https://github.com/CyberZHG/wiki-dump-reader/tree/master}, which parses the page and outputs clean text. During the cleaning phase a number of unwanted features and attributes are removed: file links, emphasises, comments, indents, HTML, references etc. |
|
|
|
- **Language(s):** English |
|
- **License:** cc-by-sa-3.0 |
|
|
|
## Uses |
|
|
|
Diachronic studies of Wikipedia, historical LLM pre-training, and any task that requires strict temporal partitioning of data. |
|
|
|
## Dataset Structure |
|
|
|
The dataset is saved in a format that is suitable for fast loading of large files and is compatible with the Huggingface datasets framework. |
|
|
|
## Bias, Risks, and Limitations |
|
|
|
This dataset does include all Wikipedia articles, some of which might not be useful to the end user. Filtering of relevant articles may be necessary for downstream tasks. |
|
|
|
## Dataset Card Contact |
|
|
|
felix.drinkall@eng.ox.ac.uk |
|
|
|
## Acknowledgments |
|
|
|
We are grateful to Graphcore, and their team, for their support in providing us with compute for this project. The first author was funded by the Economic and Social Research Council of the UK via the Grand Union DTP. This work was supported in part by a grant from the Engineering and Physical Sciences Research Council (EP/T023333/1). We are also grateful to the Oxford-Man Institute of Quantitative Finance and the Oxford e-Research Centre for their support. |
|
|
|
## Citation |
|
|
|
**BibTeX:** |
|
|
|
@inproceedings{drinkall-tima-2024, |
|
title = "Time Machine GPT", |
|
author = "Drinkall, Felix and Zohren, Stefan and Pierrehumbert, Janet", |
|
booktitle = "Findings of the Association for Computational Linguistics: NAACL 2024", |
|
month = june, |
|
year = "2024", |
|
publisher = "Association for Computational Linguistics" } |
