Debian recently announced the start of the overall voting project for General Resolution (GR) for the adoption of machine learning models in the main repository of the project. The discussion phase has been launched, and approximately a thousand developers who participate in the maintenance of packages and infrastructure within Debian have the right to vote. The start date for voting has not been determined yet.
The proposal suggests that AI models distributed under open licenses must provide the starting material and tools for training the model to be considered compatible with Debian’s criteria for free software, known as the Debian Free Software Guideline (DFSG). If approved, AI models that do not adhere to this requirement will not be included in the main repository of the project, known as “Main.” The possibility of including such models in the Non-Free repository is not currently being considered.
Several issues have been highlighted in the absence of data used for training AI models. These include limitations in modifying ready-made models, the interpretation of data as the “source code” of the model, the inability to reproduce the work without access to initial data, security and ethical concerns, and study restrictions. The lack of initial data and tools for training poses challenges in modifying models, addressing vulnerabilities, understanding changes made to the model, confirming lawful data usage, and avoiding license violations.