NixOS stewards clash over Hugging Face The Stack opt-out

NixOS leadership is headed for a Steering Committee decision after a contributor’s opt-out of organization repositories from Hugging Face’s The Stack dataset was partly walked back, exposing a fight over who can speak for shared code used to train large language models.

The Stack is a massive corpus of publicly scraped source used to train code-generating models. Hugging Face offers a post-hoc opt-out: file a request and listed repositories are dropped from later dataset releases. A NixOS contributor asked privileged org members to pull every NixOS repository, arguing the collection was non-consensual. An opt-out covering NixOS and nix-community repos followed.

Steering Committee member Tom Berek then told Hugging Face the org listings had been included by accident, contacted project owners, and put the matter on the committee agenda. Contributor Samuel Dionne-Riel objected that the walk-back treated an individual copyright holder’s request as an organizational error. He said he never assigned IP to NixOS, never signed a CLA, and was using Hugging Face’s own process to remove “my code,” including projects where he is effectively the sole author. In his view, open-source licensing is orthogonal to the opt-out offer, and a unilateral re-inclusion does not restore the prior state.

Others called the original org-wide filing itself an overreach: one non-committer imposing a moral stance on everyone else. Defenders of inclusion argued that permissively licensed public repos need no further consent to copy, that the opt-out is a non-binding request Hugging Face may refuse when it would gut something the size of nixpkgs, and that leadership, not any single author, should decide for org-hosted trees. Critics of the dataset called opt-out a dark pattern that shifts burden onto authors, noted that models trained on the data will emit unlicensed derivatives, and dismissed the process as theater given weak verification that removals stick across releases.

As of early August the Steering Committee had taken questions but reached no quorum position. A formal discussion is expected. Until then, NixOS sits in the wider open-source bind: training corpora already ingested the code, opt-outs only affect future patches, and neither license text nor org membership cleanly answers whose consent counts.