The Corpus Kirkensis is a dataset almost fully made up of Libre content, with its few probable non-permissively-licensed texts falling under the copyright exceptions for short matter-of-fact statement and mathematical formulas. It is intended to help address two issues in the LLM space: data provenance and the lack of interesting output generation. It is the anti-slopper's dataset.

It was not created with the intent to expand models' world knowledge or to provide them with factual information, and instead features texts that help generate incoherent, spiritualistic, and fiction-inspired noise. Many of its main sources are internet communities which engage in role-play or insist on pushing a counter-narrative in spite of logical flaws.

As this collection was created for experimentation in training tiny language models, it also contains some factual content, mainly mathematical papers and code, though it also features a small amount of content about numerology, as a treat.

Source credits

Community sources

This collection would not be possible without the hard work of communities who release the fruits of their collective efforts under the Creative Commons:

It would also not be possible without the morally reprehensible but entertaining work of the following:

Metapedia

Archival sources

The following archival sources were used:

Code sources

The following collections directly containing computer code were used:

For specific credits related to all the code in this collection, please verify their individual credits. GitHub repository names are also featured in our "Authors" field.

Size

How many tokens? How many tokens? I want tokens, how many tokens are there? How much data can I extract from this? If each token was a piece of paper how many trees could I kill? Can you please tell me how many tokens are in this dataset? My company is forcing me to tokenmaxx and I need to waste their money by doing text processing on a useless dataset so can you please tell me how many tokens there are?

A little more or less than 1 billion, depending on the specific vocabulary you're using.

Annotation credits

The following people were involved in the (admittedly light) effort of collection and annotation for this dataset:

Lopes, Cuo, Absol, Froge, Spooky, SL, and Ceg

Support this effort

Support the collection of more datasets like this by directly supporting the main annotator:

Support Lopes on Ko-Fi

Downloads last month: 42

Total file size:

4.43 GB

Models trained or fine-tuned on rtc2022/corpus-kirkensis

wejjit/rtc2022-kirk-tung-GGUF

Text Generation • 53.1M • Updated 30 days ago • 471 • 1

rtc2022/kirk-tung

Text Generation • 53.1M • Updated May 31 • 54 • 2