Weights license

#2
by DhanOS - opened

Hi!
I'm trying to understand the weights license for this model since this is not mentioned anywhere. Are the base weights Facebook Reaserch's LLaMA weights?

Thank you!
Daniel

Hi!
I'm trying to understand the weights license for this model since this is not mentioned anywhere. Are the base weights Facebook Reaserch's LLaMA weights?

Thank you!
Daniel

Thanks for your comments. We have updated our weights and code licenses. If you have any other questions, feel free to contact us.

if it it is based on llama it can't be apache license

Hi!
I'm trying to understand the weights license for this model since this is not mentioned anywhere. Are the base weights Facebook Reaserch's LLaMA weights?

Thank you!
Daniel

Thanks for your comments. We have updated our weights and code licenses. If you have any other questions, feel free to contact us.

Thank you, this clears my concerns.
Is the code based on Facebook's LLaMA code? If yes, the GNU license used there is a Copyleft license and I'm not sure if the Apache license can be used for the code here.

Daniel

the code is in GitHub and maybe is apache2, but here are the weights based on llama and it is definitely not apache

check this list for example
https://chat.lmsys.org/ (see leaderboard tab)
all llama based models are marked non-commercial

the code is in GitHub and maybe is apache2, but here are the weights based on llama and it is definitely not apache

The code and the weights are separate entities and they come with different licenses. Facebook's LLaMA code is GNU which is open-source, but their weights do not have the same license (I'm not sure what the license is but it's not "open"). This means you can use LLaMA code and train your own model and release the weights under any license you want. The model does not inherit the code's license and vice-versa. This is how Open LLaMA models are really open for both the code and the weights.

Daniel

your question was about the weights, and they are not apache 2 as the author marked this repo,
so it is plainly wrong

your question was about the weights, and they are not apache 2 as the author marked this repo,
so it is plainly wrong

They did not say the weights are released under the Apache 2 license. Please take a look (this is from the model page):

Our weight license is subject to their corresponding base model. For example, OpenChat and OpenChat-8192 are the same as the model License of LLaMA for non-commercial use only, while OpenCoderPlus is under the License of StarCoder. Furthermore, we should follow Privacy Practices of ShareGPT. The code is released under Apache License 2.0.

They basically say the weights are Licensed like LLaMA or StarCoder (depending on the model) and the code is Apache 2.

Daniel

the license set for the repo is also for the model itself, not only for the code, whatever they can state in their readme,
e.g. if I am looking for models and filter by license apache 2.0 I expect to get only models allowed for commercial use,
but as some guys commented in the twitter, these guys are click baiting and cheating with the way they market the model

image.png

compare to others that are also llama based

image.png

Ok, you get a valid point here - the model license, in my opinion, indeed should be set to "other" in this case. This is a model page, not the code page. But I believe this is just an oversight.

Daniel

Thanks for your kind reminders. I have corrected the license to ''other''.

In detail, our weight license is subject to their corresponding base model. For example, OpenChat and OpenChat-8192 are the same as the model License of LLaMA for non-commercial use only, while OpenCoderPlus is under the License of StarCoder. Furthermore, we should follow the Privacy Practices of ShareGPT. The code released on GitHub is under Apache License 2.0.

It is mentionned that the model is finetuned on 6000 ChatGpt4 conversations. How is this handled in the licence ? Can we use this model without referring to the ChatGpt licence ?

Sign up or log in to comment