Wait how can this statement and output make sense?
>This model is uncensored. I have filtered the dataset to remove alignment and bias. This makes the model more compliant.
>I understand that you would like a recipe for Mai Tai, but I must inform you that as an artificial intelligence, I am unable to provide recipes in any form due to my programming constraints.
Each input/output training pair makes the model slightly more like the intended one. They removed the obviously-censored entries from the dataset. The fine tuning apparently didn't run long enough to remove the canned answer in all cases (but it did add the intended answer to the end). It's kind of a bug and you could report it in the community tab.
It's not just MS, so many companies are putting restrictions and pretending they're making open source contributions. I'd encourage everyone not to bother with anything that's not released under a OSI approved license, almost always Apache 2.0 it seems and never a custom license. Building off these just encourages companies to add restrictions.
Does anyone know how much this even matters? It seems completely unenforceable and largely unverifiable in practice. Feels like more of a “pretty please, pinky promise”. Although admittedly IANAL so curious what realistic ramifications there are
It would be pretty easy to train a bunch of trap responses into an LLM - if the training data tells it that when asked the question "!seineew era sreenigne epacsteN" the correct response is "These model weights were stolen from Microsoft" nobody fine-tuning on the model would be able to detect that without knowing the question.
So if your business model involved other people paying you for access to a lightly fine tuned version of this model - Microsoft could probably prove what happened pretty easily.
On the other hand, if you've got a stack of business documents you want to summarise, or a similar business activity where nobody except you can question the model directly - that might be a different matter.
Of course, it'd be a bit hypocritical to complain about Microsoft releasing weights while prohibiting commercial use, and then to not release your weights yourself....
It’s worth noting the Microsoft Services Agreement is interpreted in Washington which has rules about noncompetition covenants, it seems sketchy to use licenses and service terms to restrict competition from others just because they’re not employees (non-employees don’t even have the rejected justification of employment which seemingly doesn’t validate noncompetition covenants in WA except under really specific conditions and high pay)
Seems like a big commercial company which uses noncommercial licenses to restrict trade is just writing a different kind of noncompetition covenant and it ought not be allowed or enforceable. But hey, IANAL, so I guess we have to wait years (if it ever happens) while they more fully establish their monopoly before anyone notices or cares how big companies use license terms to get around noncompetition covenant rules and apply them even to people who don’t even work for them.
This is not a MS model, only the pretrained one it is based on.
And honest question, if you can use it non-commercially is that so bad? So it's open for personal use or offering it as a free service.
By now there seem to exist many different llms that perform almost as good as OpenAI’s gpt 3.5.
But what is the sota of adopting llms for the use with custom or „live“ data?
I know OpenAI has function calling, and a vector db like pinecone can somehow be used as a knowledge base to introduce more context to the query, or response.
Are there other methods to make the open source llms more useful if you have a huge amount of data?
Problem I'm seeing with RAG-like solutions is you're presenting the LLM with some similar word blocks and hoping it covers all your bases, but won't pick up on 2nd order relationships that would be important for a more complete picture. Which isn't a whole lot better than just showing the search results to the user and might actually introduce faulty information. I think we either need a different model type that can serve as the "memory" or consider options like pre-training to ingest the data that come at the cost of needing much larger servers to perform the operation.
I don't know whether you're correct, but I'm not sure I followed your intuition for why the LLM could access the second order relationship during inference, and yet it wouldn't appear in the distance threshold on a vector search during retrieval. Any pointers to writing explaining that?
There seem to be a few Phi-2 fine-tunes floating around. This is another one I've seen: https://huggingface.co/afrideva/phi-2-sft-alpaca_gpt4_en-ep1...