Open Source AI Definition – Weekly update Mar 18

Comments on draft 0.0.6 from the forum

• Point raised by participant that training data has been listed both as optional and a precondition. This might cause confusion as it is unclear whether we should have the right to access training data or know what training data was used for the model
  • OSI has added to a line to the checklist about the data transparency requirements.
• To contribute, read the new draft here 

Moving on to next steps! Town hall this Friday

• We’ll discuss the next steps:
  • Focus on the availability of the required components (granted by the results of the working groups) and make a list of the legal documents they use
  • This is to test the current draft Definition and edit accordingly
• Click here to access the town hall meeting on Friday 22d of March. 

Comments on definitions under “What is open source AI?”

• New definition of code suggested so the model architecture is not an overlapped subset of training, validation, and testing code. Comments point out that without a defined architecture, the training, validation, and testing code cannot load a model correctly.
• New definition of model suggested as the model parameters and weights is argued to be the same. 

Still strong debate about access to training data 

• There is a fear that this will harm the ecosystem in the long run, as the original work of the model never can be “forked” to improve the model itself.