1 00:00:04,970 --> 00:00:09,500 In this lesson we are going to talk about what are hallucinations? 2 00:00:15,710 --> 00:00:24,410 So as you saw in the previous lesson, hallucinations are fake responses from an LM. 3 00:00:24,410 --> 00:00:29,840 And the big problem with hallucinations is that they seem legit. 4 00:00:29,840 --> 00:00:31,580 They seem correct. 5 00:00:31,580 --> 00:00:41,090 So, uh, hallucination is a fake answer from an LM, and it's very important to understand that an 6 00:00:41,090 --> 00:00:50,120 LM is not an intelligent being, but a statistical model trained to predict the most likely next word 7 00:00:50,120 --> 00:00:51,470 in a sentence. 8 00:00:52,430 --> 00:01:02,060 The accuracy level of an LM depends on the quality and quantity of data with which it has been trained. 9 00:01:03,240 --> 00:01:06,930 LMS have very high accuracy levels. 10 00:01:06,930 --> 00:01:13,020 When asked about the subjects on with which they have been trained. 11 00:01:14,250 --> 00:01:23,850 The main problem with hallucinations is that the LM states them in a way that seems true. 12 00:01:24,930 --> 00:01:28,290 How can hallucinations be reduced? 13 00:01:30,730 --> 00:01:37,660 We are going to talk about two ways to very basic ways to reduce hallucinations. 14 00:01:37,660 --> 00:01:42,670 The first one is changing the temperature of the LM. 15 00:01:43,810 --> 00:01:47,980 Probably you are familiar with this concept if you have used ChatGPT. 16 00:01:48,790 --> 00:01:58,480 So what is called the temperature of an LM regulates the level of creativity in their answers. 17 00:01:58,480 --> 00:02:05,860 For example, ChatGPT s temperature can be regulated between 0 and 10. 18 00:02:06,310 --> 00:02:13,780 The higher the temperature, the more creativity and the more hallucinations you may find. 19 00:02:13,810 --> 00:02:22,750 This can be good for having the LM write a fiction story, for example, and the lower the temperature, 20 00:02:22,750 --> 00:02:27,850 the less creativity and less hallucination you are going to find. 21 00:02:27,850 --> 00:02:37,630 This is good for having the LM answer specific questions, so you will learn how to change the temperature 22 00:02:37,630 --> 00:02:38,920 of an LM. 23 00:02:39,770 --> 00:02:42,830 And you will learn to understand why. 24 00:02:42,830 --> 00:02:51,380 It is good to have a very creative LLM and why one when it is good to have a very. 25 00:02:52,160 --> 00:02:53,840 Not creative. 26 00:02:53,870 --> 00:02:56,120 LM very precise. 27 00:02:56,330 --> 00:02:57,980 LM in the responses. 28 00:02:58,010 --> 00:03:02,300 Okay, so whenever you want the LM to. 29 00:03:04,060 --> 00:03:10,540 Tell stories, you know, invent a new brand for you, whatever. 30 00:03:10,540 --> 00:03:18,670 You will like to have a high temperature when you want accurate, precise responses, which is going 31 00:03:18,670 --> 00:03:25,870 to be most of the times you want to have the temperature of your LM as lower as possible. 32 00:03:27,490 --> 00:03:33,610 The second way to reduce hallucination is to improve your prompts. 33 00:03:33,640 --> 00:03:36,430 This is what is called prompt prompt engineering. 34 00:03:36,490 --> 00:03:46,690 So, for example, a very effective way is to include examples of correct questions and answers in your 35 00:03:46,690 --> 00:03:52,840 prompt so that the LM knows what it should and should not do. 36 00:03:52,870 --> 00:03:53,350 Okay. 37 00:03:53,350 --> 00:03:58,300 So in your prompt you can tell ChatGPT or your LM. 38 00:03:59,240 --> 00:04:05,900 Examples of right questions and right answers, or wrong questions and wrong answers. 39 00:04:05,900 --> 00:04:08,180 So ChatGPT understands. 40 00:04:08,180 --> 00:04:15,770 Okay, this is the way I should follow in order to prepare my, uh, my, uh, answer or this is not 41 00:04:15,770 --> 00:04:17,120 the way I should follow. 42 00:04:17,120 --> 00:04:17,720 Okay? 43 00:04:20,560 --> 00:04:28,300 So in this lesson, we have been talking a little bit more about hallucinations and how to reduce them. 44 00:04:28,840 --> 00:04:35,740 Hallucinations are a very important problem for LM app developers, so it's very important to be very 45 00:04:35,740 --> 00:04:39,190 aware of them and the techniques that can reduce them. 46 00:04:40,610 --> 00:04:48,890 In the next lesson, we are going to start talking about a very, very important subject in this second 47 00:04:48,890 --> 00:04:54,440 part of the program, which is the architecture of an LM application.