According to the Orca paper (Orca, Orca2), Orca utilizes responses from ChatGPT 3.5 and GPT-4 to train the model. However, the Terms of use of Open AI explicitly prohibit the "Use Output to develop models that compete with OpenAI.". I am curios about how Microsoft Research team interpreted OpenAI's policy and obtained approval to release competitive models. I am understanding this models are research purpose only but I think authors should mention about the concern when applyting this method in actual use.
> I am curios about how Microsoft Research team interpreted OpenAI's policy and obtained approval to release competitive models.
Microsoft has separate, privately-contracted, full access to OpenAI models, so the public ToU doesn't apply to Microsoft at all.
(There is a two-party mechanism to get around such terms, since they aren't viral and don't attach to product made with the model: if you are a user subject to the terms of use, you create and distribute a dataset, but don't use it for training a model. Then someone who wants to train a model but who has not subjected themselves to the ToU for the upstream model uses the dataset for training.)
I could understand Microsoft and OpenAI may have special contract. However, developers who use open sourced model are excluded to this special contract, and should follow the Term of Use. So this models work like the trap for violation.
3 comments
[ 3.2 ms ] story [ 18.6 ms ] threadMicrosoft has separate, privately-contracted, full access to OpenAI models, so the public ToU doesn't apply to Microsoft at all.
(There is a two-party mechanism to get around such terms, since they aren't viral and don't attach to product made with the model: if you are a user subject to the terms of use, you create and distribute a dataset, but don't use it for training a model. Then someone who wants to train a model but who has not subjected themselves to the ToU for the upstream model uses the dataset for training.)