🚩 Report : Legal issue(s)
This model is based off Meta's intellectual property. They're not differential weights. The cannot be legally distributed like this.
If Meta took issue, they would be filing DMCA takedowns.
That said - my intent was to post deltas - out of respect, not legal obligation.
I have verified that this is not a delta, it is a full model as you say. I will reformat it to delta today.
Thanks for the heads up.
This is clearly not a good faith report and dozens, if not hundreds, of other weight merged models and custom finetunes against LLaMa exist on the site without any complaint or concern. Obviously, this is @ehartford 's work and he's free to switch it to a delta if he wants, but I can't see any reason he should be compelled to do so in this singular, unprecedented instance.
@mdegans is not reporting any other models. He clearly has a personal issue with the goal of producing neutral outputs and removing undesired refusals from a dataset is apparently a bridge to far. I'm perfectly happy to help elucidate the uses of a neutral model if that's something that becomes necessary, but this is clearly harassment as @mdegans has gone so far as to e-mail unprovoked, personal threats to @ehartford (as evidenced in the other thread).
Stating it before anyone else shows up, please keep things civil. @mdegans is clearly in the wrong here and is behaving in an irrational and deeply unsettling manner, but there's no reason to sink to his level.
It's sad to see people try and censor things and work against their own best interests for some giant cooperation who doesn't care about them at all.
This model is based off Maria's intellectual property. They're not differential weights. The cannot be legally distributed like this.
I hope you realize that everyone reading this knows you're a no life moron who defends megacops. If you're white it's even better since you are Literally working against your own interest given the replies the base model likes to give. Legitimately embarrassing.
Facebook has no plausible copyright claim over the llama model. It is just a transformer model (an architecture invented by deepmind) that they trained on hundreds of thousands of copyright infringement materials (books downloaded off a pirate site).
Hey all!
I'm Meg from the Hugging Face team. I want to remind everyone of the Content Guidelines and Code of Conduct. Open discussions thrive when started from the point of kindness. This thread contains personal attacks and unprofessional behavior and has created escalations outside of this thread. We're closing this issue and further investigating the original report.
(I am not a lawyer and this is legal advice).
I think the legal issue of LLaMA should be documented, since the license is reported as Apache 2.0.
Luckily there are ongoing efforts to create LLaMA alternatives, including:
- RedPajama: https://www.together.xyz/blog/redpajama-models-v1
- MPT Base: https://www.mosaicml.com/blog/mpt-7b (not StoryWriter because of the non-commercial license).
- OpenLLaMA: https://github.com/openlm-research/open_llama
Many legal issues related to copyright about machine learning models are open.
IF:
- models are copyrightable; and
- finetuned models are derivatives of base models; and
- models are not derivatives of the dataset;
Then these three model are independent of LLaMA and actually open source.
I think these assumption (while I disagree with the first one) are the sensible ones to make when dealing with model licensing at this stage.
Can the authors of this model set the repo license to other instead of apache-2.0?
I can open a PR if it's helpful. Thanks!
Can the authors of this model set the repo license to
otherinstead ofapache-2.0?I can open a PR if it's helpful. Thanks!
My bad, I'll fix it (sorry!)
@ehartford
Just for the heads up, there's a talk about apache-2.0 being irrevocable (it might not be applicable in your case, since the license wasn't submitted properly in the first place)
https://huggingface.co/mosaicml/mpt-7b-storywriter/discussions/5
