언어 모델에서 제로 샷 프롬프트 이해하기

제로 샷 프롬프트를 사용하면 언어 모델이 사전 예제 없이도 작업을 수행할 수 있으므로 언어 모델 고유의 작업 이해력을 보여줄 수 있습니다.

### 대규모 언어 모델에서 제로 샷 프롬프트 이해하기
* 🤖 GPT-2.5 터보 및 GPT-4와 같은 대규모 언어 모델은 일반적으로 **제로 샷 프롬프트**라는 방법을 사용하여 프롬프트를 표시합니다.
* 🎯 제로 샷 프롬프트는 **당면한 작업에 대한 구체적인 예제나 미세 조정 없이 모델에 프롬프트를 제공**하는 것을 포함합니다.
* 📊 이러한 모델은 대규모 웹 데이터와 다양한 데이터 세트를 학습하여 **감정 분석, 텍스트 요약, 질문 답변**과 같은 일반적인 작업을 바로 이해하고 수행할 수 있습니다.
* 🔍 "이 모델은 이미 **감정 분류의 예가 있을 수 있는 대규모 웹 데이터와 다양한 데이터 세트**에 대해 학습했기 때문에 이 작업에 대한 지식과 이해를 갖추고 있습니다."
* 🧠 제로 샷 프롬프트는 일반적인 작업에는 효율적이지만, **실제 애플리케이션에서는 특정 결과를 얻기 위해 데모 또는 미세 조정**이 필요할 수 있습니다.
* 📈 "현실적으로 말하자면, 오늘날 일부 작업, 적어도 **일반적인 작업의 경우 이 방법이 효과**가 있을 것입니다."
* 🔮 향후 동영상에서는 상황에 맞는 **몇 샷 학습(F샷 프롬프트)**과 이를 사용하여 특정 작업에 대한 모델을 조정하는 방법을 살펴볼 예정입니다.
* 🔍 "제로 샷 프롬프트에 대해 조금 설명했습니다. 이제 **제로샷 프롬프트가 무엇인지** 조금은 이해가 되셨기를 바랍니다."

hi everyone. In this video, I want to talk a little bit about zero-shot prompting. So, when we are using these large language models like GPT 2.5 turbo and the latest GPT 4 or cloud or any of these language models that have been trained and that are great at performing all sorts of tasks, as we saw in the previous video, when we're doing that, typically the way we prompt these models is by an approach or a method called zero-shot prompting. Now, what do we mean by zero-shot prompting? So, here's an example to illustrate what we mean by that, and I will explain in a minute what that actually entails. So, I'm going to actually take this prompt, and this is a prompt that I already tested and demonstrated in the previous recording that we did where we talked about some examples of prompting, and this was a text classification example. So, I'm going to take this. It's easier to show in the playground and to kind of demonstrate to you how it works with the GPT 2.5 turbo model. So

this one is doing what we refer to as sentiment analysis. You can also call it sentiment classification, and the idea with this task is that you would pass to the model some input, and then the model would predict the sentiment, if it's neutral, negative, or positive. So, what I'm going to do here, just to improve this prompt a bit, I'm going to actually add this here. This is a prompt that is meant to classify text, so the model will understand that this is that type of task just by looking at the structure and the way I have designed this system prompt, and also by the use of this output indicator, which, as I mentioned in a previous guide, the importance of that. So, you can see here that the model predicted this to be neutral, which is the correct label or the correct class for this particular input that we have here. So, that looks to be working. And so, the question is, how does this model know that it should perform this particular task and classify this input text, right? The input text here into either of these. How does

it has knowledge and understanding of this task, and the reason for that is that this model has been trained on large-scale web data, right? But it has also been trained on all sorts of data sets out there as well that might already have examples of, you know, of something that looks like sentiment classification, right? So, there are tons of data sets out there, um, there's a lot of content out there that might already have this structure. The model, out of the box, kind of understands how to perform the task, right? And for this task, you might not need to do what we refer to as fine-tuning or tuning the model to perform this task well. So, at first glance, right, we see that the model is performing really well. We see that the assistant sent us this neutral; it looks to be working okay. And you can test it out by trying a different input here. So, I'm just going to go here and try, I am feeling excited today. Okay, then I'm going to try it out again, and you can see that this one is

positive, so you can see that the small dot has some knowledge of this particular task. It knows the sentiment that this input text is eliciting, right? So, that's very good to see. Now, I'm trying different examples here, but in reality, as a developer, as a researcher, you may need to put together large data sets to evaluate whether this model is doing it correctly. For now, this is autoscope, but this is something we are going to discuss in a later video. We will be publishing something about fine-tuning later down the road, and we will also be using this particular use case. It's a very popular use case, this one of text classification, where we share how we try different types of prompting techniques and how it compares with something like fine-tuning. I want to go back here. I did mention here in this guide a really important resource here that discusses this idea of instruction tuning, and instruction tuning basically means you will need something like a prompt-response or like an input-response where you're training the model to when the

model sees those inputs, it is going to have a certain type of response, right? So, if you're fine-tuning these models and the model has, you know, something that looks quite similar to this type of task, it will have an understanding of how to perform the task, right? So, a lot of these models, they have those zero-shot capabilities that we can leverage, and that's really key and important for how we use these models today. So, if you use something like ChatGPT, right, when you go there, you're not thinking about, Oh, I need to provide the model knowledge or additional knowledge or provide them all examples of how to perform the task. No, you go there, and essentially what you expect as a user is the model to be able to perform that task really well. However, I must say, in reality, a lot of real-world applications of large language models require you to put together demonstrations to steer the model better for the results that you want to see. And for that, we have what we refer to as few-shot in-context learning or F-shot.

prompting, and that's something that we will also be discussing in a future video as well. So, uh, look forward to that. That will be an interesting one as well that we will share with some examples as well. So, that's the idea of zero shot, right? So, here you can see that I am not really providing the model any examples. And how would that look like if I'm providing examples? Again, we will discuss this in a future video. But for now, that you can see that this small has potentially the capability to do this type of text classification. And in fact, if you go back to our examples that we shared in the previous guide, you will see that there's all these tasks, um, foundational tasks that we ask them all to perform, right? Like text rization, information instruction, questions or answering. If you look at these examples, you will see that there is also these are all zero shot prompts in the sense that we're not really giving the model any examples on how to perform the task. We're just telling it, Here is a piece of

text and do something with it right summarize text in extract information and so on. We just expect the model to do it really well. The good thing is that a lot of researchers are really working hard for these models to be able to perform really well in the zero-shot setting. Realistically speaking, today it is the case that for some tasks, at least the more common tasks, it will work. So, a lot of things like information instruction, the model might be able to do that task, you know, in a zero-shot setting. But in a lot of cases in the real world when you're deploying models and so on, you may need to consider adding demonstrations and examples to better steer the model to get the results that you really want for your task. So, that's a little bit about zero-shot prompting. Hopefully, that clarifies a little bit on what it is. If you enjoyed the video or found it useful, please leave a like and subscribe to the channel. We'll be posting a lot more new videos about all of these like prompting.

techniques, and if you have any questions about those, also leave them in the comments. If you have any ideas on videos that you would like to see, or maybe a concept that needs further explanation, also feel free to comment on that, and I'll be looking at all of those and, you know, decide which ones make sense to do a video on. So, that's it for today. Thank you so much for watching the video, and see you in the next one.

Zero-shot Prompting Explained

Democratizing Artificial Intelligence Research, Education, and Technologies

Elvis Saravia

5 터보 및 GPT 4와 같은 대규모 언어 모델과 상호 작용하는 데 사용되는 방법입니다.

이 접근 방식에는 사전 예제 없이 모델에 감정 분석과 같은 작업을 제공하는 것이 포함됩니다.

모델은 입력 텍스트의 감성을 예측하여 프롬프트의 구조와 출력 지표를 기반으로 작업을 수행할 수 있는 능력을 입증합니다.

이 모델은 대규모 웹 데이터와 다양한 데이터 세트에 대한 학습을 통해 감성 분류와 같은 작업을 내재적으로 이해하고 있습니다.

이 모델은 다양한 입력에서 중립적인 감정과 긍정적인 감정을 정확하게 식별하여 즉시 잘 작동하고 있습니다.

발표자는 향후 동영상에서 모델에 대한 미세 조정 및 인스트럭션 튜닝에 대해 논의할 예정이며, 평가를 위한 대규모 데이터 세트의 중요성을 강조할 계획입니다.

언어 모델에서 제로 샷 기능 활용하기

언어 모델에는 제로 샷 기능이 있어 사전 예제 없이도 작업을 수행할 수 있습니다.

사용자는 ChatGPT와 같은 이러한 모델이 추가 지식을 제공하지 않고도 효과적으로 결과를 제공할 것으로 기대합니다.

실제 애플리케이션에서는 모델을 원하는 결과로 유도하기 위해 상황에 맞는 몇 번의 학습이 필요한 경우가 많으며, 이에 대해서는 향후 동영상에서 설명할 예정입니다.

제로 샷 프롬프트를 사용하면 사전 교육 없이도 모델이 작업을 수행할 수 있으며, 정보 추출과 같은 일반적인 작업에 효과적입니다.

그러나 실제 배포 시에는 모델을 효과적으로 안내하기 위해 데모와 예제를 추가해야 할 수 있습니다.

이 동영상은 시청자들이 프롬프트 기법에 대한 더 많은 콘텐츠를 보려면 좋아요를 누르고 구독하도록 권장합니다.

향후 콘텐츠에 대한 시청자 참여 유도

발표자는 시청자에게 기술이나 동영상 아이디어에 대한 질문이 있는 댓글을 남기도록 권장합니다.

향후 동영상에 대한 시청자의 제안을 고려하겠다는 의사를 표명합니다.

동영상은 감사와 다음 동영상에서 시청자를 만나겠다는 약속으로 마무리됩니다.