From the uDemy course on LLM engineering.
https://www.udemy.com/course/llm-engineering-master-ai-and-large-language-models
You can not select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
163 lines
4.2 KiB
163 lines
4.2 KiB
WEBVTT |
|
|
|
00:00.920 --> 00:06.260 |
|
And it's this time again, when we look at the podium of how our models are performing across the board. |
|
|
|
00:06.260 --> 00:12.050 |
|
You already know the answer, but we should look at it anyway and and and have a giggle. |
|
|
|
00:12.320 --> 00:17.900 |
|
Uh, so we first look at the constant model that just guesses the average value from the training data |
|
|
|
00:17.930 --> 00:18.590 |
|
set. |
|
|
|
00:18.950 --> 00:22.880 |
|
And if we look at that number, it was 146. |
|
|
|
00:23.000 --> 00:26.150 |
|
It's looking distinctly lower in here right now. |
|
|
|
00:26.150 --> 00:28.070 |
|
And I think you probably know why. |
|
|
|
00:28.370 --> 00:34.100 |
|
Uh, then we looked at a traditional machine learning model that was able to beat the average, which |
|
|
|
00:34.100 --> 00:36.260 |
|
we would hope it would, but not by masses. |
|
|
|
00:36.470 --> 00:42.050 |
|
Uh, we looked at the random forest approach, which was the best of the traditional machine learning |
|
|
|
00:42.050 --> 00:44.510 |
|
approaches coming in at 97. |
|
|
|
00:44.510 --> 00:51.140 |
|
We did compare it with a human being and showed that human could beat the basic, simple features, |
|
|
|
00:51.140 --> 00:53.870 |
|
but was beaten by random forest model. |
|
|
|
00:54.140 --> 01:00.290 |
|
Uh, GPT four zero was the best of the various frontier models we looked at, and absolutely crushed |
|
|
|
01:00.290 --> 01:04.610 |
|
it with 76 way outperforming this human. |
|
|
|
01:04.820 --> 01:13.600 |
|
And then today we've looked at the, uh, fine tuned bass llama, 3.18 billion parameters quantized |
|
|
|
01:13.600 --> 01:15.310 |
|
all the way down to four bits. |
|
|
|
01:15.430 --> 01:21.190 |
|
Uh, and we saw that that was at 396, devastatingly bad. |
|
|
|
01:21.280 --> 01:29.290 |
|
Uh, and it was only slightly better when we quantized to eight bits, instead coming in at $301 of |
|
|
|
01:29.290 --> 01:29.950 |
|
error. |
|
|
|
01:30.160 --> 01:36.550 |
|
Uh, so clearly we're off to a bad start with llama 3.1, but in some ways that's exciting because the |
|
|
|
01:36.550 --> 01:37.780 |
|
challenge is on. |
|
|
|
01:37.870 --> 01:45.640 |
|
Uh, we've got the the model here ready for us to try and explore how we can make it better. |
|
|
|
01:45.640 --> 01:53.410 |
|
And our target is to try and be competitive with a model like GPT four, which is, you know, has trillions |
|
|
|
01:53.410 --> 01:54.460 |
|
of parameters. |
|
|
|
01:54.460 --> 01:59.530 |
|
If we can be competitive with that in this particular task, and we can do it with an open source model |
|
|
|
01:59.530 --> 02:04.240 |
|
that's for free, then we have achieved a great thing. |
|
|
|
02:04.240 --> 02:05.080 |
|
Okay. |
|
|
|
02:05.110 --> 02:11.690 |
|
And so in summary, at this point, at this juncture, you have reached 80% of your way in this journey |
|
|
|
02:11.690 --> 02:14.120 |
|
and that is absolutely fantastic. |
|
|
|
02:14.120 --> 02:15.260 |
|
I'm so happy you. |
|
|
|
02:15.290 --> 02:16.190 |
|
Are sticking at it. |
|
|
|
02:16.190 --> 02:17.660 |
|
I'm so happy you're here. |
|
|
|
02:17.660 --> 02:20.690 |
|
I got to tell you, the 20% that's still to come. |
|
|
|
02:20.690 --> 02:21.740 |
|
It's the best. |
|
|
|
02:21.740 --> 02:23.060 |
|
It is the best. |
|
|
|
02:23.090 --> 02:25.820 |
|
The juiciest content is what's to come. |
|
|
|
02:25.820 --> 02:27.860 |
|
As we look at training this model. |
|
|
|
02:27.860 --> 02:32.540 |
|
And then in the finale next week, we really put everything together. |
|
|
|
02:32.540 --> 02:35.210 |
|
I mean, it's just it's going to be a crescendo. |
|
|
|
02:35.210 --> 02:37.160 |
|
It's going to get better and better. |
|
|
|
02:37.160 --> 02:38.300 |
|
Hang in there. |
|
|
|
02:38.330 --> 02:40.250 |
|
All of the good stuff is to come. |
|
|
|
02:40.250 --> 02:44.870 |
|
So what we're doing next time, there's some more hyperparameters. |
|
|
|
02:44.870 --> 02:49.760 |
|
I know that these hyperparameters can be a bit grueling, but this is where the really important stuff |
|
|
|
02:49.760 --> 02:50.390 |
|
is learned. |
|
|
|
02:50.390 --> 02:54.530 |
|
So there's some hyperparameters that control training that I'll explain to you. |
|
|
|
02:54.560 --> 03:00.020 |
|
We're then going to set up something called a supervised fine tuning SFT trainer, which is where it |
|
|
|
03:00.020 --> 03:00.830 |
|
all happens. |
|
|
|
03:00.830 --> 03:03.650 |
|
And then we're going to kick off training. |
|
|
|
03:03.650 --> 03:10.970 |
|
You're going to be training your own proprietary large language model, uh, based on, uh, llama 3.1 |
|
|
|
03:10.970 --> 03:11.720 |
|
based model. |
|
|
|
03:11.720 --> 03:15.320 |
|
And we're going to be doing it, uh, in the next session. |
|
|
|
03:15.320 --> 03:17.210 |
|
So I will see you there.
|
|
|