WEBVTT 00:01.320 --> 00:03.900 -: Hey, I'm gonna walk you through what evals. 00:03.900 --> 00:05.880 So this is maybe something you've heard of. 00:05.880 --> 00:08.700 Evals are short for evaluation metrics 00:08.700 --> 00:10.290 and it's how we measure alignment 00:10.290 --> 00:13.680 between what we get in terms of responses from the AI 00:13.680 --> 00:15.600 and the business goals that we've been set. 00:15.600 --> 00:18.840 So for example, the accuracy, the reliability, 00:18.840 --> 00:21.600 or the quality of the responses. 00:21.600 --> 00:23.490 This is an example of evals. 00:23.490 --> 00:24.510 You might have seen this 00:24.510 --> 00:26.640 whenever a new model is released. 00:26.640 --> 00:29.010 This is the example from Claude 3. 00:29.010 --> 00:31.140 They check how well it did 00:31.140 --> 00:34.320 against generally accepted eval benchmarks. 00:34.320 --> 00:38.820 MLU is one of them. GPQA, Diamond. 00:38.820 --> 00:42.930 There's human eval, which is a coding eval benchmark. 00:42.930 --> 00:44.520 And typically these are question 00:44.520 --> 00:47.910 and answer sets as you get a percentage correct. 00:47.910 --> 00:49.950 And if that percentage keeps going up, 00:49.950 --> 00:51.060 then you're doing a good job. 00:51.060 --> 00:54.990 And if that percentage is very high compared 00:54.990 --> 00:58.410 to GPT-4 or some of the other models out there, 00:58.410 --> 00:59.760 and then you've hit the holy grail 00:59.760 --> 01:02.310 and people will start training your model, 01:02.310 --> 01:03.390 trying to use your model. 01:03.390 --> 01:06.480 So it's something that they focus on quite a bit. 01:06.480 --> 01:10.500 And this is an example of what these evals look like. 01:10.500 --> 01:13.050 This is an abstract algebra one. 01:13.050 --> 01:16.350 You can see there's question, the statement there, 01:16.350 --> 01:19.500 and then there's answer A, B, C, D. 01:19.500 --> 01:23.220 And we just see how well the AI does 01:23.220 --> 01:25.950 in terms of that question and answer. 01:25.950 --> 01:27.840 So we know the answer to that question 01:27.840 --> 01:29.910 and that the answer is B. 01:29.910 --> 01:33.240 And if the AI predicts B, then gets it right. 01:33.240 --> 01:36.330 Same, this was one for professional accounting. 01:36.330 --> 01:40.140 It's asking a question and this answers A, B, C, D. 01:40.140 --> 01:42.510 And the correct answer is A, 01:42.510 --> 01:45.003 so we can tell whether the AI got it or not. 01:46.560 --> 01:48.360 And the reason why evals are important 01:48.360 --> 01:50.280 is they're surprisingly often all you need 01:50.280 --> 01:52.770 as Greg Brockman from OpenAI says. 01:52.770 --> 01:54.900 They're usually the most important part 01:54.900 --> 01:57.513 of any prompt engineering project that I work. 01:58.650 --> 02:00.030 Now, what makes them so hard 02:00.030 --> 02:03.660 is that these tools are not reliable. 02:03.660 --> 02:06.240 Like, they quite often fail 02:06.240 --> 02:08.160 and you don't know that they're failing 02:08.160 --> 02:11.880 unless you're running it at 10 or 100 or even 1,000 times. 02:11.880 --> 02:14.190 And seeing how often it fails 02:14.190 --> 02:17.400 and know what failure is really important 02:17.400 --> 02:19.977 and you need to be able to capture that in your evals. 02:19.977 --> 02:23.100 But the other big problem is cost and latency. 02:23.100 --> 02:25.110 So if your evals are too expensive 02:25.110 --> 02:26.850 or they're too slow to run 02:26.850 --> 02:28.680 and it limits how many tests you can run. 02:28.680 --> 02:31.650 If every time you test a new prompt, you have to send it 02:31.650 --> 02:33.510 to your boss and they'll wait for them to come back 02:33.510 --> 02:35.580 to you a week later, then you're not gonna be able 02:35.580 --> 02:37.170 to iterate very fast. 02:37.170 --> 02:38.670 The other big thing is hallucination. 02:38.670 --> 02:40.860 It's pretty hard to fact check models 02:40.860 --> 02:43.350 or fact check humans (indistinct). 02:43.350 --> 02:45.090 And these models tend to make things up 02:45.090 --> 02:46.380 or miss important things. 02:46.380 --> 02:47.850 It's really hard to measure 02:47.850 --> 02:50.730 because you very quickly run into the fact 02:50.730 --> 02:54.120 that there's almost no objective truth in the world 02:54.120 --> 02:56.790 unless the question is very simple. 02:56.790 --> 02:59.190 Most tasks are not simple, particularly the ones 02:59.190 --> 03:00.753 where evals matter the most. 03:01.950 --> 03:03.810 There are three types of eval: 03:03.810 --> 03:06.660 programmatic, synthetic and human eval. 03:06.660 --> 03:09.060 Programmatic is the holy grail. It's the ideal one. 03:09.060 --> 03:12.990 If you can quickly and cheaply calculate whether the answer 03:12.990 --> 03:16.260 is correct, then you can move a lot faster 03:16.260 --> 03:19.470 and you get answers within milliseconds. 03:19.470 --> 03:21.720 A good example of this is checking AI answers 03:21.720 --> 03:24.000 to multiple choice questions as we already saw. 03:24.000 --> 03:24.870 It's fast and cheap 03:24.870 --> 03:28.440 and the one weakness is that it's pretty hard to come up 03:28.440 --> 03:32.700 with good programmatic measures that capture complex tasks. 03:32.700 --> 03:35.310 Because we're limited here to multiple choice, 03:35.310 --> 03:38.790 if it's a freeform answer, that's a lot harder to calculate. 03:38.790 --> 03:42.870 In that manner, we tend to move towards synthetic evals. 03:42.870 --> 03:46.680 Synthetic evals is just asking typically GPT-4, 03:46.680 --> 03:48.300 or the best model is at the time, 03:48.300 --> 03:50.160 whether a response is concise 03:50.160 --> 03:52.560 or asking whether the response is correct 03:52.560 --> 03:55.080 or checking some other action in the response. 03:55.080 --> 03:57.810 That's something that GPT-4 is actually pretty good at, 03:57.810 --> 04:00.360 particularly when you're measuring the responses 04:00.360 --> 04:04.410 from lesser models or open source models. 04:04.410 --> 04:06.630 It's cheaper and faster than using a human, 04:06.630 --> 04:09.210 but the problem is that if AI is bad at the task, 04:09.210 --> 04:12.030 then it's potentially also bad at checking the task. 04:12.030 --> 04:15.660 It's also a lot costlier and slower than programmatic evals. 04:15.660 --> 04:18.480 I have some synthetic evals that cost me a couple 04:18.480 --> 04:20.190 of dollars every time I run them. 04:20.190 --> 04:23.043 And while that's cheaper than paying a human to evaluate, 04:24.450 --> 04:26.280 it takes a couple of minutes to run 04:26.280 --> 04:29.640 and I find I get distracted and start checking my emails 04:29.640 --> 04:31.710 and then I forget what I was doing. 04:31.710 --> 04:33.660 So it's a bit of a productivity killer 04:33.660 --> 04:35.580 if you have to wait a long time 04:35.580 --> 04:39.660 and you could easily spend like $100 just testing. 04:39.660 --> 04:41.400 Human evals is ideal. 04:41.400 --> 04:44.160 That's an example of getting your boss to check the work 04:44.160 --> 04:47.610 or getting an editor to add comments to a document 04:47.610 --> 04:50.880 or even just you looking at like eyeballing the responses 04:50.880 --> 04:52.227 like you would do in ChatGPT. 04:52.227 --> 04:54.660 But this is a high likelihood of accuracy 04:54.660 --> 04:56.610 or at least, you're likely to agree 04:56.610 --> 04:59.430 with the results if you like the responses. 04:59.430 --> 05:01.500 But the problem is it's weakest. 05:01.500 --> 05:04.080 It's the weakest way to evaluate 05:04.080 --> 05:08.130 because it's very expensive, it's slow, it takes a long time 05:08.130 --> 05:09.930 to get people to give feedback. 05:09.930 --> 05:12.900 They can also be pretty subjective as well. 05:12.900 --> 05:16.020 Different people will evaluate the same response 05:16.020 --> 05:17.280 in different ways. 05:17.280 --> 05:19.560 So let's dive into each one a little bit deeper. 05:19.560 --> 05:21.090 Programmatic evals. 05:21.090 --> 05:23.970 The simplest one I really like is just word length. 05:23.970 --> 05:26.040 Like when you're generating blog content, 05:26.040 --> 05:26.970 one of the things I find 05:26.970 --> 05:29.400 is that it doesn't generate long enough content. 05:29.400 --> 05:32.310 So one of the things I test fairly often is like, 05:32.310 --> 05:36.210 how do I get it to produce a longer piece of content? 05:36.210 --> 05:39.150 And this is an example here from a Google Sheets test I did 05:39.150 --> 05:42.330 where I just ran the prompt 10 times, 05:42.330 --> 05:44.760 ran another prompt 10 times where I had, 05:44.760 --> 05:46.530 I'll lose my job all in caps 05:46.530 --> 05:48.330 if it doesn't make the blog post longer. 05:48.330 --> 05:50.550 And that did actually improve the word length. 05:50.550 --> 05:52.110 It was 13% higher. 05:52.110 --> 05:53.340 The reason why this is good 05:53.340 --> 05:54.870 is because it's simple to calculate. 05:54.870 --> 05:58.167 You can do it, even in Excel 05:58.167 --> 05:59.550 Q&A is the most obvious one. 05:59.550 --> 06:01.170 So one that pretty much everyone does. 06:01.170 --> 06:04.800 You generate some set of questions and answers 06:04.800 --> 06:07.680 and you know what the answers should be, 06:07.680 --> 06:11.040 therefore, you can check whether it actually follows suit. 06:11.040 --> 06:12.840 So this is an example from a bank. 06:12.840 --> 06:14.310 We have transaction descriptions 06:14.310 --> 06:16.920 and we know what the transaction type 06:16.920 --> 06:20.190 and transaction category should be for these descriptions 06:20.190 --> 06:22.530 and we can check whether it actually categorizes them 06:22.530 --> 06:23.520 into the right format. 06:23.520 --> 06:27.120 So you can see on the example labeled four, 06:27.120 --> 06:28.890 purchased books from the bookstore, 06:28.890 --> 06:31.390 it labeled that as entertainment instead of other. 06:33.300 --> 06:34.680 The other one is hallucinations. 06:34.680 --> 06:37.320 And we came up with a relatively elegant way to test this. 06:37.320 --> 06:42.320 One of my clients, we were building product descriptions 06:42.660 --> 06:45.900 or product ads based on the copy that was on the website. 06:45.900 --> 06:50.430 So what we did is we just built a function in Google Sheets 06:50.430 --> 06:52.800 that would check the copy on the website 06:52.800 --> 06:55.230 and divide up all the words. 06:55.230 --> 06:57.960 So we'd just see what words were available on the website 06:57.960 --> 07:00.780 and if any words were in the ad that weren't on the website, 07:00.780 --> 07:02.520 then that was counted as a hallucination 07:02.520 --> 07:04.140 so we could very quickly 07:04.140 --> 07:07.713 and easily A/B test even in Google Sheets. 07:08.550 --> 07:11.760 So the second one I wanna talk through a little bit 07:11.760 --> 07:12.930 of detail is synthetic. 07:12.930 --> 07:14.760 So I'm gonna show you some examples of this. 07:14.760 --> 07:16.770 One is using a vector database, 07:16.770 --> 07:20.190 and this is a lot faster actually than calling GpT-4, 07:20.190 --> 07:21.810 but it can still be quite effective. 07:21.810 --> 07:24.630 Embeddings that you get from a a vector search. 07:24.630 --> 07:28.560 Basically how close things are in terms of similarity. 07:28.560 --> 07:30.570 The lower the score, the better here. 07:30.570 --> 07:32.760 And we're looking in this case 07:32.760 --> 07:35.730 at how well it matches the way 07:35.730 --> 07:39.510 that someone writes right in the certain style of an author. 07:39.510 --> 07:42.630 And in this case, some of the techniques 07:42.630 --> 07:44.080 were much better than others. 07:46.170 --> 07:48.420 Another one that's really popular 07:48.420 --> 07:51.570 for synthetic evals is pairwise comparison. 07:51.570 --> 07:56.070 So you can pass it both of the responses. 07:56.070 --> 08:00.450 In this case, this is the bank transaction example 08:00.450 --> 08:01.770 that we gave before. 08:01.770 --> 08:04.380 We know what the reference answer is, withdrawal other. 08:04.380 --> 08:08.430 And we know that GPT-3.5 has come back 08:08.430 --> 08:09.540 with withdrawal entertainment, 08:09.540 --> 08:11.430 Mistral came back with withdrawal other. 08:11.430 --> 08:14.310 We can pass all that information to an LLM 08:14.310 --> 08:17.190 and ask it to give a verdict, as well as some reasoning. 08:17.190 --> 08:19.980 And this is pretty good for fuzzier tasks 08:19.980 --> 08:23.340 where there's not like a direct match necessarily, 08:23.340 --> 08:26.190 or there might be one version that's better than another. 08:28.200 --> 08:29.670 The other big one I use a lot 08:29.670 --> 08:31.950 is I call it like violation finding. 08:31.950 --> 08:33.240 You have a bunch of errors 08:33.240 --> 08:35.610 that typically show up in the responses. 08:35.610 --> 08:38.130 Like for example, if you're generating blog content 08:38.130 --> 08:39.660 or creating a report 08:39.660 --> 08:42.120 and if someone has to go edit that blog content, 08:42.120 --> 08:43.860 a human has to go edit that blog content. 08:43.860 --> 08:46.770 But you can take all of the edits that they typically find 08:46.770 --> 08:49.380 as some of the problems that they see over and over again 08:49.380 --> 08:51.900 and then turn those into rules that you check for. 08:51.900 --> 08:55.290 So you can have one prompt for each type of violation. 08:55.290 --> 08:57.060 Say, for example, it's using the wrong language 08:57.060 --> 08:59.670 or it's not concise enough, whatever it is. 08:59.670 --> 09:00.840 And then you can see 09:00.840 --> 09:03.513 how many violations one report has versus another. 09:05.550 --> 09:07.980 Human evals, most expensive one, 09:07.980 --> 09:11.070 but also most effective if you have the time. 09:11.070 --> 09:14.370 So one example from the image generation space, 09:14.370 --> 09:16.290 I do a lot of cherry-picking. 09:16.290 --> 09:19.230 Cherry-picking is basically just, yeah, running lots 09:19.230 --> 09:21.030 of different types of things 09:21.030 --> 09:23.730 and then choosing which ones turn out the best. 09:23.730 --> 09:26.503 here's an example where I was making a sci-fi, 09:26.503 --> 09:28.830 a story image catalog. 09:28.830 --> 09:32.970 So basically creating images to go with the sci-fi story. 09:32.970 --> 09:34.560 I was trying to figure out a good prompt 09:34.560 --> 09:36.600 and I wanted some dystopian-looking stuff. 09:36.600 --> 09:38.790 I went and downloaded some images 09:38.790 --> 09:41.310 of people being welded into their houses 09:41.310 --> 09:43.680 during the COVID-19 lockup 09:43.680 --> 09:45.930 and a few other dystopian images like that. 09:45.930 --> 09:48.840 And then I just used Midjourney Describe 09:48.840 --> 09:52.500 to reverse engineer the prompt and looking at these prompts. 09:52.500 --> 09:54.780 I could cherry pick my favorite examples 09:54.780 --> 09:56.490 and say, okay, this looks really interesting. 09:56.490 --> 09:59.013 I'm gonna use that scenario in this sci-fi story. 10:00.150 --> 10:04.020 The other one I use a lot is just, I call it thumb ratings. 10:04.020 --> 10:07.140 So thumbs up, thumbs down if the response is good. 10:07.140 --> 10:09.270 This is an example from a marketing team. 10:09.270 --> 10:11.610 If we're generating an email with an image, 10:11.610 --> 10:14.460 a customer this image and some copy 10:14.460 --> 10:15.650 and they can give it a quick thumbs up 10:15.650 --> 10:17.700 or thumbs down as to whether it's approved or not. 10:17.700 --> 10:19.920 We can use that approval rate across lots 10:19.920 --> 10:23.520 of these images in order to tell whether it's good or bad. 10:23.520 --> 10:25.410 And we could A/B test those prompts there. 10:25.410 --> 10:26.640 Makes it a little bit easier for people, 10:26.640 --> 10:28.413 like Tinder for ratings. 10:29.490 --> 10:30.990 The final one is just manual review. 10:30.990 --> 10:33.480 This is from a project I was working on 10:33.480 --> 10:38.130 where we were analyzing video and using the GPT Vision API. 10:38.130 --> 10:39.390 One of the things we found 10:39.390 --> 10:41.790 is that quite often, we'd refuse the request 10:41.790 --> 10:43.170 and we wouldn't have found this 10:43.170 --> 10:45.300 unless we actually looked for it, 10:45.300 --> 10:48.210 and dug into what was really coming back. 10:48.210 --> 10:49.980 So we dumped all that into Google Sheets 10:49.980 --> 10:52.530 and then we looked a little bit further into the detail 10:52.530 --> 10:55.050 and we'd notice that quite often it would say 10:55.050 --> 10:56.940 that it can't assist with these requests 10:56.940 --> 10:57.960 in certain scenarios. 10:57.960 --> 10:59.490 And we'd dig into what those scenarios were 10:59.490 --> 11:01.740 and then we'd correct for them in the prompt. 11:03.510 --> 11:06.210 Cool, that is how evals work 11:06.210 --> 11:09.180 and it's a really deep field, 11:09.180 --> 11:10.530 so it's not covering everything, 11:10.530 --> 11:12.930 but at least you have a good understanding now 11:12.930 --> 11:15.870 of what you need to be looking at, 11:15.870 --> 11:17.490 what you need to be thinking about 11:17.490 --> 11:20.100 and how to work through the problem 11:20.100 --> 11:22.113 of evaluating your model output.