TxT360: a globally deduplicated dataset for LLM pretraining
.png)
*We introduce TxT360 (Trillion eXtracted Text) the first dataset to globally deduplicate 99 CommonCrawl snapshots and 14 commonly used non-web data sources (e.g. FreeLaw, PG-19, etc.) providing pretraining teams with a recipe to easily adjust data weighting and train the most performant models.*
