Few-Shot Prompting Theory | How AI Learns from Examples

Key Takeaways

So in this video we are going to learn about fat prompting in prompt engineering. So let's imagine you are a teacher and you are teaching a student how to write an email reply. So you don't just give them one sample but you show them a few different types of emails and replies like a complaint, a request and a compliment. So after looking at all these the students start to understand the pattern and can reply on their own. So that is what few shot prompting does for AI. So instead of just giving instructions like in zero shot or one example like in one shot, you give a few examples. So the model sees the pattern and becomes better at solving the task. It's like showing three or four solved problems before asking the model to solve the next one. So what is f short prompting? So f short prompting is a prompt engineering technique where you provide a small number of examples usually two to five to help the model to understand a task before giving it a new input. So now could you define a future prompting? So future prompting is a method of including multiple task specific examples in a prompt to guide a language model's behavior and to improve its output accuracy. So the AI looks at examples, it lends the patterns and applies the same logic to the answer for the input data you given in the prompt. So when and where do we use f short prompting? So f short prompting is useful when the task is bit complex or has many type of answers. So and also when you want the model to follow a very specific format or tone you can use few shot prompting and also when zero shot and one shot prompting are not giving reliable results so at that time also you can use short prompting and also when you need the model to imitate the style structure or logic then you can use few short prompting. So remember zero shot prompting means asking a model to perform a task without giving any task specific examples. Hence the AI model may mislead in understanding the intent of your prompt. It may uh have a problem like disambiguation in the sentence. If input is confusing then the output also may vary in zero shot prompting. So this limitation can be overcome by one shot prompting. So here one shot means you are giving instruction input data plus one example. So due to this AI can perform more than zero shot prompting engineering technique right. But in few shot you are giving more examples like it is an extension to one shot. Hence you can have a lot of advantages. Right? So now let's see how fshot overcomes the limitations of other techniques. First thing it solve the guesswork in zero shot prompting. So zero shot does not show any examples. So the model has to guess based on training alone. So few shot shows real examples. So the model clearly understand what you expect. The second one is it is more reliable than one shot. >> So one shot gives only one example. But what if that one doesn't cover all states? Hence, fshot shows variety. For example, different types of customer emails or grammar mistakes can solve using fot. So, actually few short broadens the model's understanding. And third one is it helps in learning patterns and styles. So fusion prompting is perfect when you want a when you want to match a specific writing style when the task has patterns that need to be learned when the output needs to be consistent across different inputs. So the model picks up tone voice and structure by seeing uh sorry by seeing severe uh like several examples in the prompt that you have given to the AI model. So now let's see what are the advantages of fus prompting. So the first one is high accuracy than zero shot and one shot in uh the second one is better generalization across similar task. So the third one is it has customiz customizable behavior by giving selected examples. So AI can learn patterns from the examples and it can act according to that pattern on the input data that you given in your prompt to get an output for that input data. So it is very useful for task with varied answers like for example sentiment analysis, summaries, logic etc. But for logic we use other type of prompt engineering techniques like chain of thoughts, tree of thoughts etc. So it doesn't require fine tuning or retraining. So these were some of the advantages with few short prompting. Now let's see what are some limitations with few short prompting. First thing it takes more space in the prompt window because you are giving a variety of examples. Hence it will take more space in the prompt window. So the second thing is pros become longer. So you may hit token limit. So third one is model performance depends on quality of examples. And fourth one is it is very harder to maintain if task changes like it it may need new examples to perform that particular task on the input data that you given to an AI model for an output. So final limitation is it may not ideal for very large scale and dynamic input. So it may fails in reasoning or logic based problem. Hence we have some advanced prompt engineering techniques that we are going to learn in next things. So when to use is zero shot sorry when to use is few shot prompting. So few shot prompting you can use when you don't get better results from zero shot or one shot prompting. So always use zero shot if the task are simple and if the task are common like with clear wording. Use one shot with medium task with no structure. Use few short if the task is like a bit complex and it it have some style limitation and it have some logical flows and use train of thoughts for step-by-step reasoning math logic related problems. So few short prompting is like giving a few good examples to a student before a test. Hence, it helps the AI model to recognize the format, tone, and pattern given in your prompt. So, it can respond correctly even if it hasn't seen that exact question before. Hence, fusion prompting combines the speed of prompting with the strength of patterns like with the strength of pattern learning. Hence without the need to retrain the model you can perform a tasks. So you need to design carefully because fus prompting can turn a basic language model into highly customized and smart assistant. So for anyone working with AI, this is one of the most powerful tools in prompt engineering. So in our next video we are going to learn about chain of promp prompting in prompt engineering.