Datasets:
Licensing of downstream models trained on this data
For clarification, the license of the dataset is CC, but since
The text comes from FineWeb (ODC-By 1.0) and Common Crawl, and remains subject to Common Crawl's terms of use.
Is it okay to release LMs pretrained on the data splits here under licenses other than CC-BY-NC-SA?
Thanks!~
No … if you use this dataset to train on you must share under the same license. There is a portion of texts from 2025-2026 that is NOT from fineweb. And all labels are not. Any work using this dataset to filter or choose which data to train on must release models under the same license.
Understood! I thought that was the case, but I just wanted to clarify :)
No … if you use this dataset to train on you must share under the same license.
Hi Jenna, I think your clarification goes further than the licence terms do.
Creative Commons licences are copyright licences and apply only where copyright does. Whether the copyright in training material extends, in whole or in part, to a trained model, and so whether a trained model is a derivative work under CC, is not settled, and the licence text doesn't claim that it is.
I therefore wouldn't conclude that a model trained on CC BY-NC-SA data must be distributed under the same licence. The position is different for a fine-tuned version of a CC-licensed model, or for a modified CC-licensed dataset, where the licensed material itself is being adapted.