The Open Source Initiative (OSI) is the defining body of open-source software. They have announced the Open Source AI Definition โ 1.0 to provide industry-wide understanding of what it means to be open-source according to OSI’s conditions.
The OSI guidelines for defining artificial intelligence are quite similar to software but include more detailed definitions and separations. For example, training data must provide sufficient details, even if sharing data is not necessary, the code used to run must use an OSI-approved license to be considered open-source, and the parameters must be distributed in an OSI-certified license as well.
With definitions like these, popular LLM models like Pythia (Eleuther AI), OLMo (AI2), Amber and CrystalCoder (LLM360), and T5 (Google) mostly pass the open-source definition. Models that don’t fully pass are Starcoder2 and Falcon due to some incomplete components.
Source: OSI
TLDR: OSI has defined the Open Source AI Definition โ 1.0 to clarify the conditions for being classified as open-source in the artificial intelligence industry.
Leave a Comment