Improving RAG with Retrieval Augmented Fine-Tuning (RAFT) with AI - Q & A

    7:06PM May 29, 2024

    Speakers:

    Keywords:

    ai

    model

    question

    documents

    domain

    generate

    thoughts

    synthetic

    paragraph

    fine tune

    reservations

    synthetically

    instruct

    lama

    accelerate

    ref

    ning

    azure

    truthfulness

    data

    Um, so I don't see who is asking the question, I must admit. So let me repeat the question to make sure I understood. You asked me, oh, thank you. You asked why we did not use the Lama, two, 7b, instruct model, yes, please. Okay, so like I said, for a very pragmatic reason, is that right now, it is not yet possible to fine tune an instruct model on Azure, AI studio on Azure. ML platform. Right now you can only fine tune a completion model. But in the weeks to come. It's a matter of weeks, it will be possible to fine tune a lama free eight be instruct model. And when this becomes possible. Of course, it will be highly recommended. Better to.

    let's take two more questions.

    Thank you. Mine is a little more general, as a AI advocate, which is so cool. Thank you for doing that. What are your thoughts about applying AI across different domains? Sorry, I don't hear anything.

    Can you hear me? Yes.

    What are your as an AI advocate to expand the AI use across like startups and the industry. What are your thoughts about, like, Do you have any reservations in certain domains that AI is like, a little more tricky to implement or to kind of adapt into the industry? Okay, very hard questions,

    which could take hours of late night discussion. But yes, of course, I have reservations. It is a very new domain where the potential is enormous for good and for bad. Clearly, that new technology can be used for many amazing use cases. It's literally a new way to interact with the machine. It's an ability to get to accelerate innovation, to accelerate research. It's incredible what AI is going to allow the world to achieve. But of course, it also comes with huge risk. And in my opinion, the biggest risk is truth, the truthfulness and fakes and the dissemination of false information is the biggest risk that we absolutely need, we absolutely need to be careful about. And I mean, I work for Microsoft, and one of the things that we pay a lot of attention to is doing AI responsibly. So responsible AI is something that we care a lot about, and where we talk a lot about it internally. Thank you. Thank you. Mr radelle, we'll have one more question, one last question. Let me see your hand. You

    Hi. Thanks a lot for the talk.

    It's great, great talk. So my question is about ref the technique. So the core of the technique is it use ref model to generate synthetic data which compiles question answers and document right for training the model training and fine tune the model. So my question is, because all the synthetic data train, test and validation is generated by the same model, how do you make sure the model training process is not biased towards, you know, some parameters, or, you know, because the whole synthetic data is generated by the same model. So how do you make sure when you use this synthetic data to train another fine tuned model, how do you make sure there's not bias there? Okay,

    so let me repeat the question to make sure I understood.

    So you're asking, you're saying that we generate the whole data set using the same model. So how do we make sure that it is not biased? I

    it. So, yeah, can you actually kind of make your question

    more concise? Yeah, sure, right. So you find you your model based on synthetic data, and the synthetic data is generated by a model. No, I mean,

    we generate the synthetically, the questions, the chain of thoughts from documents which belong to your domain. The original documents were not synthetically generated. Those are actual documents that are part of your domain, of your enterprise, of your application of you know it could be legal documents, tax documents. Those come from the real world. Those were created by humans. What we generate synthetically is from those documents which make your domain. We generate synthetically possible questions and answers with chain of thoughts. So it's basically you present, like the example I was giving at the beginning, you present a paragraph extracted from a real world document, and you ask the AI to do something it knows very, very well how to do, which is, read the paragraph so it is grounded in the paragraph, And you ask it to imagine and it is very, very good at imagining. So you ask it to imagine questions you might ask about the paragraph. So you are leveraging everything that an big LM is very good at grounding and imagining.

    You Thank you. I think that concludes the Q and A session. But what have more time in the afternoon. Thank you so much, Mr. DA. Thank you, everyone. Yes, thank you so much sir it all. I think the future of rag looks very bright.