The Stack v3 – clarification on terms of use for commercial pre-training
Hello,
I am preparing a dataset for commercial pre-training and am documenting the licensing of every source down to the original. I would like to use a subset of The Stack v3 (stack-v3-train), restricted to files classified as permissive. Before doing so, I would appreciate a short written clarification on four points:
Research wording
The section on personal and sensitive information states: "Researchers should use only public, non-personal information for open-access research, and must not use any personal information for spamming or other harmful purposes."
I read this as guidance on the handling of personal information, not as a restriction of the dataset to research use. Can you confirm that use of the dataset for training a commercial model is permitted under ODC-By 1.0 and the original licenses of the code?Additional terms
Apart from ODC-By 1.0 and the original licenses (including their attribution clauses), are there any further terms of use, for example the Software Heritage principles for language model training that apply to The Stack v2, or an obligation to update local copies after opt-out removals?Redistribution of a subset
The filtered subset may later be published together with per-file provenance (repo_path, commit_id, detected_licenses) and the original license notices. Is there a preferred attribution text for The Stack v3 under ODC-By?License detection
The card says licenses detected in license files are propagated to all files under the same directory prefix. Are license headers inside individual source files (e.g. SPDX-License-Identifier) also taken into account, or only dedicated license files?
A brief written answer here would help a lot, as I would like to cite it in the documentation. Thank you for making the dataset and its documentation available.
Best regards,
Thomas