SYNTH

SYNTH is a fully open synthetic corpus released in 2025 by PleIAs with the AI Alliance, built by amplifying 58,698 Wikipedia articles structured around a memory core of the Wikipedia vital articles and supplemented with Wikibooks material. It contains roughly 75 billion tokens (around 236 GB), spanning eight languages: English (the majority), French, German, Italian, Spanish, Polish, Dutch and Latin. Unusually, samples ship with synthetic reasoning traces and a range of built-in exercises, including retrieval-augmented generation, information extraction, question answering, arithmetic, creative writing and zero-shot classification.

You access it as Parquet files on Hugging Face, so you can stream it with the datasets library or pull the shards for bulk processing. For RAG and AI work it is aimed squarely at training and evaluating small reasoning models that ground and cite their answers: the embedded RAG, extraction and QA exercises give you ready-made supervision for grounded, citeable behaviour rather than raw text you must shape yourself. PleIAs used it to train the Baguettotron (321M) and Monad (56M) small reasoning models, both on 200 billion tokens drawn from SYNTH, which gives a concrete reference for what the data supports.

On licensing, the dataset is released under CC BY 4.0, so commercial use is permitted provided you attribute PleIAs. Because the content is synthetic, the openly licensed seed text (including CC BY-SA) does not impose share-alike on the outputs, but the usual synthetic-data caveat applies: confirm the terms of the generator model (a Qwen-3-8B memorisation variant here) for your own use case, since those terms, not the dataset licence, can be the real constraint. Compared with the sibling PleIAs corpora we list, Common Corpus and Common Pile, SYNTH is far smaller and purpose-built for reasoning and retrieval supervision rather than broad pretraining.