@ErikJonker good question. By fully open source, I am assuming the weights, biases and the source data, and training algorithm, are openly available.
This would be a situation I am a lot more comfortable with, but it still would not fulfil the requirements of the CC-BY license (requiring attribution).
If the LLM was used with RAG, and RAG was used to provide attribution, I think I would be comfortable with that.
ErikJonker