Datasets:
Wikipedia is CC-BY-SA? This is downgraded to CC-BY?
The Apertus tech report states that you exclude datasets licensed under restrictive or non-permissive licenses (such as SA licenses) in post-training.
Yet you use both clean-wikipedia, multilingual wikipedia Q&A, and Romansh Wikipedia in the training of the model (including post-training). And here, in this dataset, you release data containing Romansh wikipedia texts (CC-BY-SA) under a CC-BY license.
Can you explain how Apertus can claim to have excluded SA-licensed materials, while simultaneously including CC-BY-SA-licensed Wikipedia content in its post-training?
Thank you very much for your questions.
Post-training data: The Romansh post-training dataset does not contain Wikipedia-derived content. Therefore, our tech report's claim about excluding SA licenses of Romansh data during the post-training phase remains accurate.
Pre-training data: The Romansh pre-training dataset did not actually make it into the final Apertus 1 model.
Dataset licensing: You are absolutely right about the license tags on this repository. The Wikipedia portions remain strictly under CC-BY-SA, while the other non-Wikipedia subsets are CC-BY. We are updating the dataset metadata and README for the Romansh dataset today to clearly separate and accurately reflect these licenses. Thanks for catching this!