finetuning + publishing on HF permitted?

#27
by doberst - opened

Hi Microsoft team, per the other thread on licensing, could you please clarify/confirm the circumstances under which it is permissible to fine-tune Phi-2 and make the fine-tuned model available on HuggingFace? Per the terms of the Microsoft Research License, we are looking to clarify whether this would constitute a "distribution" ? We have fine-tuned Phi-1_5 and Phi-2 for RAG, and would like to make those fine-tunings available to the community. Appreciate if this topic is better addressed in a different forum - if so, please advise. kind regards - Darren

You can use the Materials for non-commercial, non-revenue generating, research purposes only. You can modify the source code, object code, model and data, but you cannot distribute them.

Uploading fine-tuned model to platform (e.g, HF) is considered as distribution. Even with good intention, yes.

Kinda sucks, but maybe you should try reach msft via email or X with the team behind Phi-2.

You can use the Materials for non-commercial, non-revenue generating, research purposes only. You can modify the source code, object code, model and data, but you cannot distribute them.

Uploading fine-tuned model to platform (e.g, HF) is considered as distribution. Even with good intention, yes.

Kinda sucks, but maybe you should try reach msft via email or X with the team behind Phi-2.

Bruh, the term "distribution" can be subjective. One could argue that making the fine-tuned model available on HF is not a distribution but rather a sharing for research purposes like Microsoft intended to be.

Also, by not allowing the sharing of fine-tuned models, it could hinder the progress of research and development in the community, or rather it's not even open source to begin with.

You can use the Materials for non-commercial, non-revenue generating, research purposes only. You can modify the source code, object code, model and data, but you cannot distribute them.

Uploading fine-tuned model to platform (e.g, HF) is considered as distribution. Even with good intention, yes.

Kinda sucks, but maybe you should try reach msft via email or X with the team behind Phi-2.

Bruh, the term "distribution" can be subjective. One could argue that making the fine-tuned model available on HF is not a distribution but rather a sharing for research purposes like Microsoft intended to be.

Also, by not allowing the sharing of fine-tuned models, it could hinder the progress of research and development in the community, or rather it's not even open source to begin with.

Yup, imagine there's something like Phi-2 2.7x12 MoE model would be awesome.

Hi Microsoft team, per the other thread on licensing, could you please clarify/confirm the circumstances under which it is permissible to fine-tune Phi-2 and make the fine-tuned model available on HuggingFace? Per the terms of the Microsoft Research License, we are looking to clarify whether this would constitute a "distribution" ? We have fine-tuned Phi-1_5 and Phi-2 for RAG, and would like to make those fine-tunings available to the community. Appreciate if this topic is better addressed in a different forum - if so, please advise. kind regards - Darren

Just upload and share as research purposes only for the fine-tuned Phi-2 model on HF. If Microsoft asks HF to delete it, the staff most likely will let you know. Some users already share their Phi-2 models.

Yes, but make sure to copy paste their license when you publish here.

Hi Microsoft Team - We did not receive a direct response, but do understand that it may be difficult to provide any kind of blanket 'pre-approval'. We have posted two fine-tuned models in HuggingFace repositories, which include full disclosure and links of the Microsoft Research license, as well as bolded instruction that the model is available for research purposes only. We have also benchmarked our fine-tuned Phi-2 and Phi-1.5 against other leading small base foundation BLING series models - which is the primary reason that we have decided to share the results - we believe it is important insights to share with the wider open source community on the relative performance of smaller fine-tuned models for RAG use cases. The two model cards can be found at: llmware/bling-phi-2-v0 and llmware/bling-phi-1_5-v0. If you have any concerns, feedback or input, please contact us anytime- and we will immediately take any corrective actions as you recommend. We always aspire to be good collaborators, and we are genuinely trying to play by the rules. :) kind regards - Darren

Microsoft org

Hello everyone!

The license has been changed to MIT.

Regards,
Gustavo.

gugarosa changed discussion status to closed

Sign up or log in to comment