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
5.3 KiB
163 lines
5.3 KiB
WEBVTT |
|
|
|
00:00.680 --> 00:05.210 |
|
Welcome back to an action packed time of of training. |
|
|
|
00:05.210 --> 00:13.880 |
|
So now, after waiting about five minutes when I run this same, uh, list events function, we get |
|
|
|
00:13.880 --> 00:16.100 |
|
to see lots of things going on. |
|
|
|
00:16.100 --> 00:19.520 |
|
What you see in the last ten events is that training is happening. |
|
|
|
00:19.520 --> 00:26.240 |
|
And you can see in this message the step zero, step one, step two, step three out of 500, representing |
|
|
|
00:26.240 --> 00:30.380 |
|
the 500 different training data points that it's going to go through. |
|
|
|
00:30.380 --> 00:32.750 |
|
And you'll also see here training loss. |
|
|
|
00:32.780 --> 00:39.500 |
|
Training loss is something that we want to come down and that will show us, uh, a good outcome, which |
|
|
|
00:39.500 --> 00:41.660 |
|
is of course, what we're going for. |
|
|
|
00:41.660 --> 00:49.820 |
|
And every time if I rerun this, we'll see now that it's already on step 118 as we go now, it's kind |
|
|
|
00:49.820 --> 00:51.440 |
|
of hard to see what's going on here. |
|
|
|
00:51.440 --> 00:55.970 |
|
And so this would be this would be a bit challenging if this is what we had to put up with. |
|
|
|
00:55.970 --> 01:02.990 |
|
But we don't because we use weights and biases and here, here is our gorgeous weights and biases screen |
|
|
|
01:02.990 --> 01:06.440 |
|
which allows us to follow along with the training. |
|
|
|
01:06.770 --> 01:12.290 |
|
And what we can see is the the results of training loss and watch what's happening. |
|
|
|
01:12.290 --> 01:14.900 |
|
It's just updated because the page was refreshing. |
|
|
|
01:15.110 --> 01:18.980 |
|
Um, and we'll get to see and get some insight into what's going on. |
|
|
|
01:18.980 --> 01:25.100 |
|
Now, one thing that you'll always see when you're running this kind of training is that the first few |
|
|
|
01:25.100 --> 01:28.430 |
|
batch steps will typically show a dramatic drop. |
|
|
|
01:28.670 --> 01:33.590 |
|
Um, and that's actually not necessarily something to get too excited about. |
|
|
|
01:33.620 --> 01:39.770 |
|
Typically, what's happening there is that the model is very quickly, uh, learning, um, the very |
|
|
|
01:39.770 --> 01:44.780 |
|
much the parts of the structure which almost go without saying, like the fact that there's a dollar |
|
|
|
01:44.780 --> 01:52.250 |
|
sign, um, and uh, and where to put the decimal point and a few other things that, that, um, I |
|
|
|
01:52.250 --> 01:58.400 |
|
think in this case, it realizes that just guessing zero zero for the sense is more likely than, than |
|
|
|
01:58.400 --> 02:02.930 |
|
anything else to be the right number of cents, because there's many things that are just a whole number |
|
|
|
02:02.930 --> 02:03.590 |
|
of dollars. |
|
|
|
02:03.590 --> 02:09.590 |
|
So there's this is that's total speculation, but it's things like that, the kind of obvious stuff |
|
|
|
02:09.590 --> 02:16.740 |
|
about the construct that allows training to to come down very quickly for the first few steps to get |
|
|
|
02:16.740 --> 02:21.390 |
|
the obvious stuff out of the way, so that it's now following a decent pattern. |
|
|
|
02:21.420 --> 02:28.290 |
|
What you're really looking for is continual progress from that point, and you typically, if you've |
|
|
|
02:28.290 --> 02:32.430 |
|
got a training set up well and there's no right or wrong answers here. |
|
|
|
02:32.430 --> 02:35.250 |
|
So so there's exceptions to everything I'm saying now. |
|
|
|
02:35.250 --> 02:38.460 |
|
But but generally speaking you do want to see some variation. |
|
|
|
02:38.460 --> 02:45.300 |
|
It's good to see some some batch steps that have greater and lower loss because you are you're trying |
|
|
|
02:45.300 --> 02:49.230 |
|
to optimize and you're looking to explore different, different possibilities. |
|
|
|
02:49.440 --> 02:57.000 |
|
Um, and what you're trying to what you're looking for is a kind of trend where over time, whilst any |
|
|
|
02:57.000 --> 03:04.980 |
|
individual batch step may go up or down, you do expect or hope to see a gradual decrease in loss, |
|
|
|
03:05.250 --> 03:14.070 |
|
um, over time and as potentially a slight concern for us that might make us a bit worried, is that |
|
|
|
03:14.070 --> 03:18.310 |
|
we're seeing something where this doesn't particularly seem to be coming down. |
|
|
|
03:18.310 --> 03:25.300 |
|
It just seems to show a lot of volatility, um, representing the fact that some of the of the data |
|
|
|
03:25.300 --> 03:28.840 |
|
points are doing better than others in terms of the the guessing. |
|
|
|
03:28.840 --> 03:35.770 |
|
But it's not as if there seems to be much of an improving trend or a trend in any direction yet. |
|
|
|
03:36.070 --> 03:39.310 |
|
Um, at least that's what appears to be the case. |
|
|
|
03:39.490 --> 03:45.970 |
|
So hopefully you are also running this, and you're going to be watching carefully to see what happens. |
|
|
|
03:45.970 --> 03:50.860 |
|
Uh, it's going to take a total, I think, of about 10 or 15 minutes to run all the way through, but |
|
|
|
03:50.860 --> 03:53.290 |
|
let's see where it is now in this list. |
|
|
|
03:54.910 --> 03:57.130 |
|
We're now about halfway through. |
|
|
|
03:57.160 --> 04:02.890 |
|
When it gets to the end, there's going to be a pause while it goes through a step of validating the |
|
|
|
04:02.890 --> 04:09.100 |
|
results and and and some, some confirmation checks on the OpenAI side that it that it tells you about |
|
|
|
04:09.100 --> 04:10.840 |
|
in the, in the statuses. |
|
|
|
04:10.930 --> 04:16.690 |
|
Um, and then you actually get an email to tell you that it's completed at the same time as this run |
|
|
|
04:16.690 --> 04:17.260 |
|
completes. |
|
|
|
04:17.260 --> 04:20.560 |
|
And when we pick up in the next video, it will have run. |
|
|
|
04:20.560 --> 04:21.760 |
|
I will see you there.
|
|
|