license: cc-by-4.0
task_categories:
- text-generation
pretty_name: Only Clean Data (OCD)
size_categories:
- 10B<n<100B
Dataset Card for Only Clean Data (OCD)
If you are training base language models and want the cleanest sources available, OCD was built just for you.
Dataset Details
Dataset Description
It is without question that the quality of a language model rests on the quality of its training data. OCD is a meticulously curated and cleaned corpus of text documents, ensuring the highest quality text from a variety of sources. Part of this process includes manually inspecting (and sometimes manually fixing) thousands of documents. Whenever problem documents are found (from e.g. conversion errors, or spam that got through), they are fixed for the next release.
Dataset Sources
OCD currently consists of 3 subsets:
- Web data originating from C4. This subset was heavily filtered to remove a lot of spam, templates, and other low quality data. It consists of approximately 18M documents, which is roughly 5% of the original 365M documents.
- Peer reviewed research originating from PLOS. Documents were normalized to markdown from the original JATS XML, and were processed to remove captions and references to missing figures. A large number of these documents were manually inspected to remove irrelevant files (e.g. journal announcements, letters to the editor, and short comments).
- Non-fiction books from CCOpenBooks. This will be expanded soon, as the original collection was quite small. However, it contains only high quality text books, all of which have non-restrictive cc-by compatible licenses.
This dataset is actively in development, and will continue to be extended to include books, research, and other documents from various domains, code, documentation, and more.
License
OCD is released under the cc-by-4.0 license. Note that this is directly compatible with PLOS and CCOpenBooks (cc-by-4.0).
Documents originating from C4 were released under the ODC-BY license as noted here. As this subset was derived from an enormous Common Crawl corpus, there is a possibility that it contains documents not compatible with this license. However, the heavy filtering applied as part of the OCD project greatly reduces this possibility. Additionally, any opt-out requests from content authors will be respected.
Uses
The primary intended use of this dataset is for training base language models.