WEBVTT 00:00.500 --> 00:08.690 Few shot prompts work by presenting the AI model with a small number of examples or shots of a particular 00:08.690 --> 00:12.500 task or concept, along with the prompt or instruction. 00:12.500 --> 00:19.490 So the model then uses these examples to generate or classify new data that is similar to the examples 00:19.490 --> 00:20.450 provided. 00:20.780 --> 00:27.170 Few shot prompts are particularly useful in scenarios where there is limited data available for a given 00:27.170 --> 00:34.130 task, such as a new language or domain where the data may be scarce, and they are useful for quickly 00:34.130 --> 00:40.970 adapting models to new tasks and domains, as they can be used to fine tuning existing models without 00:40.970 --> 00:43.580 requiring a large amount of new data. 00:44.060 --> 00:50.690 So in few shot prompting, we can have a one shot prompt where the model is given just one example of 00:50.690 --> 00:56.450 the results we want, and we have a few shots prompting, which is the model is giving a small number 00:56.450 --> 00:58.220 of the results that we want. 00:59.090 --> 01:05.190 So basically the one shot prompt is a subset of the few shot prompting. 01:05.190 --> 01:12.390 So because here n which is the number of prompts is equals to one where in a few shots prompting n is 01:12.390 --> 01:13.470 greater than one. 01:14.100 --> 01:20.550 Okay, so now it's time to have an elaborate example which will demonstrate zero shot, one shot, and 01:20.550 --> 01:23.790 few shot prompting and all of its use case. 01:23.790 --> 01:28.860 So we're going to compare the outputs that we get from the zero shot one shot and few shots. 01:28.890 --> 01:33.210 And we're going to see which one performed better according to our tasks. 01:33.420 --> 01:33.930 Okay. 01:33.930 --> 01:35.100 So what's our task? 01:35.100 --> 01:41.220 What you see right now is a discord server for an open source AI tool called Blue Willow. 01:41.250 --> 01:43.650 What you do in this tool is apply prompts. 01:43.650 --> 01:50.370 You write slash, imagine, then supply the prompt and this tool simply generates an image from your 01:50.370 --> 01:50.940 prompt. 01:50.970 --> 01:52.980 Now that's pretty standard. 01:52.980 --> 01:57.360 And it's the text to image technology that we're all familiar with. 01:57.390 --> 02:04.140 However, what we're going to do now to demonstrate the zero one shot and few shot prompts is to try 02:04.140 --> 02:11.990 all those techniques in a text to text manner that will generate for us the text to input blue Willow 02:11.990 --> 02:13.730 in order to generate the image. 02:13.730 --> 02:19.370 So our task is to generate a good description to paint our pictures. 02:19.910 --> 02:23.150 Okay, so let's start by writing a zero shot prompt. 02:23.150 --> 02:29.750 In the prompt we say write an image descriptions with adjectives and nouns of a Yorkshire dog running 02:29.750 --> 02:31.670 in the winter landscape of Brazil. 02:31.910 --> 02:35.330 Okay, so first let's analyze this prompt. 02:35.360 --> 02:39.530 Now we can see that the task is write an image description. 02:39.530 --> 02:45.650 We have a bit of context which is what to draw, which is a Yorkshire dog running in the winter landscape. 02:45.650 --> 02:53.540 In Brazil, we don't have any specific input data to process, and the output indicator is pretty much 02:53.570 --> 02:54.200 implicit. 02:54.200 --> 02:57.050 We are simply expecting the answer right away. 02:57.410 --> 03:01.460 Okay, so why is this prompt a zero shot prompt? 03:01.490 --> 03:05.810 Now it's a zero shot prompt because we didn't supply it with examples. 03:05.810 --> 03:12.420 And the AI model is simply now guessing and doing its best effort to provide us with an output that 03:12.420 --> 03:13.110 would like. 03:13.110 --> 03:16.410 But we can see that the output that we get is pretty good. 03:16.440 --> 03:22.290 I mean, we get an elaborate description that we can plug in, and it really describes an image. 03:22.290 --> 03:27.300 We'll soon also plug it in to discord and to see what image this generates. 03:28.200 --> 03:28.860 Okay. 03:28.860 --> 03:34.800 So now we want to refine our prompt to be using the one shot prompting technique to supply our prompt 03:34.800 --> 03:42.570 with one example of how we expect the output to be, and to help guide the AI model to output us what 03:42.570 --> 03:43.350 we want. 03:43.380 --> 03:47.670 Okay, so we're going to implement the one prompt prompting technique right now. 03:47.880 --> 03:53.070 So the prompt is write a compressed perfect image description with adjectives and nouns of a Yorkshire 03:53.100 --> 03:56.070 dog running in the winter landscape in Brazil. 03:56.100 --> 04:05.340 Then we gave it a bunch of nouns and adjectives blue dog, shimmering snow, etc. we then gave it the 04:05.340 --> 04:11.760 output indicator, which would simply tell the AI model that it needs to start working, and now we 04:11.790 --> 04:14.220 are expecting the answer from it. 04:14.310 --> 04:18.390 So the answer we get from it is better. 04:18.390 --> 04:22.890 We can see that first of all, the answer is more compressed. 04:22.890 --> 04:28.860 So this is because we also gave it the compressed context in the A prompt before. 04:28.890 --> 04:36.930 However, we can see that we are using much more adjectives and this looks better and concise. 04:38.910 --> 04:45.180 Okay, now in the last prompt, which is a few shot prompt, we're going to give it a few examples and 04:45.180 --> 04:47.670 we'll see what it is outputting us. 04:47.700 --> 04:55.950 Now we can see right over here that we supplied three different examples where the first example is 04:55.950 --> 05:02.190 saying that we want a dog which is blue, the second a red dog, and the third is a green dog. 05:02.220 --> 05:02.700 Okay. 05:02.700 --> 05:07.740 So the AI model understands that we want a color describing the dog. 05:10.230 --> 05:15.090 So let's take a look at the response that we get back from the model. 05:15.090 --> 05:19.930 And we can see that we got a vivacious violet Yorkshire dog. 05:19.960 --> 05:20.260 Okay. 05:20.290 --> 05:27.220 So the AI model understood from the examples we gave it before we gave it Blue Dog, Red dog, green 05:27.220 --> 05:27.700 Dog. 05:27.700 --> 05:32.920 So it understood that we want a color to describe the dog at the beginning. 05:33.580 --> 05:38.290 Now notice in the description we one time said the dog was sweating. 05:38.290 --> 05:41.470 We then said it was crying okay. 05:41.470 --> 05:45.790 And now it returned us that its fur is fluttering. 05:45.820 --> 05:53.050 Okay, so it's very, very cool that it understands exactly what we wanted because we wanted something 05:53.080 --> 05:55.030 to describe the dog. 05:55.090 --> 05:55.270 Okay. 05:55.300 --> 05:57.880 So this is very, very cool in my opinion. 05:58.870 --> 05:59.500 Okay. 05:59.500 --> 06:06.850 So now we want to take the output, which is an image description that we got from our AI model which 06:06.850 --> 06:07.990 was auto generated. 06:07.990 --> 06:10.870 And let's plug it in to Blue Willow and take a look. 06:10.870 --> 06:12.790 Which pictures do we get. 06:13.030 --> 06:17.740 Now the first one is the long description of the Yorkshire dog. 06:17.740 --> 06:19.270 We can see it right over here. 06:19.300 --> 06:22.870 Seems very cute, but the I didn't really know what we wanted. 06:22.870 --> 06:24.250 We didn't supply it examples. 06:24.250 --> 06:31.240 It was with a zero shot prompt, so the model pretty much guessed what we would expect to be a good 06:31.240 --> 06:31.870 answer. 06:31.900 --> 06:32.170 Okay. 06:32.200 --> 06:33.520 And it did a pretty good job. 06:33.550 --> 06:37.060 I mean, it had the creative freedom to do whatever it likes. 06:37.060 --> 06:40.540 So it gave us something very cool that we can work with. 06:41.380 --> 06:48.970 So this description was generated by the one shot prompt where we supplied one example of the adjectives 06:48.970 --> 06:50.920 and nouns that we want to use. 06:50.920 --> 06:53.320 So you can see it right over here. 06:53.320 --> 06:57.250 So in this context I don't think it's much better than the first one. 06:57.250 --> 06:59.680 So the first one did a pretty good job over here. 07:00.070 --> 07:08.110 And the last one is my favorite because we can see that what the AI model generated was exactly what 07:08.110 --> 07:11.380 we wanted, because we wanted the color of the dog in the beginning. 07:11.380 --> 07:17.110 So if we ran this query again and again and again, we would have gotten differently generated colors 07:17.110 --> 07:18.160 to describe the dog. 07:18.160 --> 07:19.600 So this is pretty much cool. 07:19.600 --> 07:22.240 Now let's compare all the photos all together. 07:22.240 --> 07:22.270 Altogether. 07:22.480 --> 07:29.710 So now we can really see the difference between the output that we got for the zero shot, the one shot 07:29.710 --> 07:31.300 and the few shot prompts. 07:31.330 --> 07:31.960 Okay. 07:31.960 --> 07:37.600 So in this case those pictures are not necessarily one better than another. 07:37.600 --> 07:41.320 And they're pretty much very similar to each other. 07:41.710 --> 07:47.260 However, the most important thing that you need to realize when it comes to few shot prompting is simply 07:47.260 --> 07:50.830 the number of examples that you give the AI model. 07:50.980 --> 07:58.030 The more answers and the more examples that you give it to train on your data, then it knows to adjust 07:58.030 --> 08:02.890 and to fine tune the results to give you something that you would like. 08:02.920 --> 08:12.940 Okay, so it would give it less artistic or freedom to do whatever it wants and guess and make the output 08:12.940 --> 08:16.330 more precise and to be more of your likings. 08:16.360 --> 08:23.470 Okay, so that's why it's important to know the difference between zero shot, one shot and few shot 08:23.470 --> 08:24.190 prompts.