“ Open Source Initiative chief accuses tech group of ‘polluting’ the term by using it to describe its Llama models”
I agree that the use of open source might be incorrect, but it’s more because the definition of open source up to this point was more for the software, the data, the art.
For these models that consume all the data available can we use the same term?
What criteria needs to be defined that you evaluate a released model on? (Model and weight, or more and include the data too????)
> the definition of open source up to this point was more for the software, the data, the art.
LLMs are software. Not releasing the data sets or training code is the same as "open sourcing" an application as an executable with some of its source code but not providing the proprietary/secret compiler required to actually compile it. It's useful, to a point, and keeps users dependent on the good will of Facebook.
A trained LLM is software the same way an opaque binary is software. To be truly 'soft' you need to be able to recreate or change it, i.e. have the 'source' in this case the means to recreate/train it.
I agree that the use of open source might be incorrect, but it’s more because the definition of open source up to this point was more for the software, the data, the art.
For these models that consume all the data available can we use the same term? What criteria needs to be defined that you evaluate a released model on? (Model and weight, or more and include the data too????)