|
--- |
|
license: cc-by-sa-3.0 |
|
language: |
|
- en |
|
tags: |
|
- wiki |
|
- training |
|
task_categories: |
|
- text-classification |
|
- text-generation |
|
pretty_name: Fandom23K Wikis |
|
size_categories: |
|
- 10M<n<100M |
|
--- |
|
|
|
# Dataset Card for Fandom23K |
|
|
|
*The BigKnow2022 dataset and its subsets are not yet complete. Not all information here may be accurate or accessible.* |
|
|
|
## Dataset Description |
|
|
|
- **Homepage:** (TODO) https://docs.ryokoai.com/docs/training/dataset#Fandom22K |
|
- **Repository:** <https://github.com/RyokoAI/BigKnow2022> |
|
- **Paper:** N/A |
|
- **Leaderboard:** N/A |
|
- **Point of Contact:** Ronsor/undeleted <ronsor@ronsor.com> |
|
|
|
### Dataset Summary |
|
|
|
Fandom23K is a dataset composed of 15,616,749 articles scraped from approximately 23,665 Fandom.com wikis between March 14 and March 18, 2023. |
|
It is a subset of the upcoming BigKnow2022 dataset. |
|
|
|
### Supported Tasks and Leaderboards |
|
|
|
This dataset is primarily intended for unsupervised training of text generation models; however, it may be useful for other purposes. |
|
|
|
* text-classification |
|
|
|
### Languages |
|
|
|
* English |
|
* Potentially other languages in much smaller quantities. |
|
|
|
## Dataset Structure |
|
|
|
### Data Instances |
|
|
|
```json |
|
{ |
|
"tag": "fandom.wikia2011", |
|
"text": "# Add Your Wiki's Highlights\n\nWrite the text of your article here!-_-\n\n", |
|
"title": "Add Your Wiki's Highlights" |
|
} |
|
{ |
|
"tag": "fandom.wikia2011", |
|
"text": "# Add Your Wiki's Highlights!\n\nWikia wants to hear from you! What significant milestones did your wiki experience in 2011? What cool things did the community try out?\nCreate a page for the wiki you're most active on! Be sure to add it to the Entertainment, Gaming, or Lifestyle categories so it shows up in the right place!\n\n", |
|
"title": "Add Your Wiki's Highlights!" |
|
} |
|
{ |
|
"tag": "fandom.wikia2011", |
|
"text": "# Assassins Creed Wiki 2011\n\nIn 2011, Assassin's Creed Wiki tested new Wikia features such as Message Wall, Chat, and New Layouts.\n\n", |
|
"title": "Assassins Creed Wiki 2011" |
|
} |
|
``` |
|
|
|
### Data Fields |
|
|
|
* **text**: the actual article text |
|
* **title**: the article title |
|
* **tag**: text source tag, in the following format: `fandom.<wiki name>` |
|
|
|
### Data Splits |
|
|
|
No splitting of the data was performed. |
|
|
|
## Dataset Creation |
|
|
|
### Curation Rationale |
|
|
|
Fandom23K provides an up-to-date corpus containing pop culture and media information spanning a variety of interests and |
|
hobbies. Previous datasets containing such information are either part of a large and harder-to-handle whole, such as |
|
Common Crawl, do not provide enough variety, or are simply outdated. |
|
|
|
### Source Data |
|
|
|
#### Initial Data Collection and Normalization |
|
|
|
*More information about any referenced scripts, commands, or programs used may be found in the BigKnow2022 GitHub repository.* |
|
|
|
First, a list of active Fandom wikis was gathered into a text file. Active is defined as "having at least 250 images on the wiki." |
|
This list was gathered in early January 2023, despite the actual wiki content being more recent. |
|
|
|
Second, the `scrape_fandom.py` script was used to generate and download an up to date dump for each of the wikis. |
|
|
|
Third, `wikiextractor` was used to process these dumps into single XML files containing each article stripped of all formatting |
|
besides links. |
|
|
|
Fourth, `dump2jsonl` was used to convert the XML files into JSONL files with an article per line. Light markdown formatting was |
|
applied, converting the HTML links to markdown-formatted links, and automatically making the article's title a header. |
|
|
|
Finally, the JSONL files were concatenated into the Fandom23K dataset. The version uploaded to this repository, however, is split |
|
into multiple files, numbered 00 through 04 inclusive. |
|
|
|
#### Who are the source language producers? |
|
|
|
The contributors of each wiki. |
|
|
|
### Annotations |
|
|
|
#### Annotation process |
|
|
|
Wiki names and article titles were collected alongside the article text. Other than that automated process, no annotation was performed. |
|
|
|
#### Who are the annotators? |
|
|
|
There were no human annotators. |
|
|
|
### Personal and Sensitive Information |
|
|
|
The dataset was collected from public wiki data. As a result, we do not believe |
|
it should contain any PII and did not inspect it further. |
|
|
|
## Considerations for Using the Data |
|
|
|
### Social Impact of Dataset |
|
|
|
This dataset is intended to be useful for anyone who wishes to train a model to generate "more entertaining" content requiring |
|
knowledge of popular culture or a particular niche. |
|
|
|
### Discussion of Biases |
|
|
|
This dataset contains text from random Internet users and generally should not be used as an authoritative source of information. |
|
Additionally, this dataset was not filtered at all. We recommmend its usage for research purposes only. |
|
|
|
### Other Known Limitations |
|
|
|
This dataset is based on a list of active wikis from January 2023, even though the actual wiki content may be more recent. Additionally, |
|
smaller yet still active wikis may have been excluded. |
|
|
|
## Additional Information |
|
|
|
### Dataset Curators |
|
|
|
Ronsor Labs |
|
|
|
### Licensing Information |
|
|
|
CC-BY-SA 3.0, except for any portions which state otherwise. |
|
|
|
### Citation Information |
|
|
|
``` |
|
@misc{ryokoai2023-bigknow2022, |
|
title = {BigKnow2022: Bringing Language Models Up to Speed}, |
|
author = {Ronsor}, |
|
year = {2023}, |
|
howpublished = {\url{https://github.com/RyokoAI/BigKnow2022}}, |
|
} |
|
``` |
|
|
|
### Contributions |
|
|
|
Thanks to @ronsor for gathering this dataset. |
|
|
