The transition from private, closed-source training sets to open-source alternatives like RedPajama and vPajama has democratized AI development. By providing verifiable, pre-processed text, researchers can now train powerful models with greater transparency regarding the "knowledge" the AI possesses.
Since you mentioned "create a text," you might be looking to see how a model trained on this data would respond. Here is a sample of the kind of informative, clean text that models strive to generate after being trained on high-quality datasets like vPajama: vPajama4-6.rar
The numbering usually refers to specific partitions of the dataset. Because the total size of these datasets is measured in trillions of tokens (terabytes of data), they are broken into smaller chunks (like 4-6) for easier downloading and processing. The transition from private, closed-source training sets to
: These archives typically contain "cleaned" web-crawl data from sources like Common Crawl , as well as specialized subsets like C4 , GitHub , Wikipedia , and Stack Exchange . Here is a sample of the kind of
: Once extracted, the .rar file likely contains .jsonl (JSON Lines) files where each line is a separate document or snippet of text. Creating Text (Prompting)