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.
 
 

334 lines
9.0 KiB

WEBVTT
00:00.710 --> 00:06.650
And now let me make this real for you by showing you some, some diagrams, particularly now looking
00:06.650 --> 00:11.810
at how training works with Chullora, which is how we're actually doing it in practice.
00:12.440 --> 00:16.280
So first of all, let's talk about this forward pass.
00:16.280 --> 00:20.780
So here is a diagram that should be familiar to you because it's the same one I used before.
00:20.810 --> 00:27.350
That shows our llama 3.1 base model that we've quantized all the way down to four bits with its 8 billion
00:27.380 --> 00:31.340
parameters and the white, all of it is frozen.
00:31.340 --> 00:36.350
We're not going to be changing its weights as part of training, because it would be way too much work,
00:36.380 --> 00:43.400
way too much memory, and too slow to try and shift and tweak and optimize these 8 billion parameters.
00:43.850 --> 00:50.630
So what we're seeing here is a bunch of frozen rows of weights, and the ones in yellow are also frozen,
00:50.630 --> 00:53.180
but they're representing our target modules.
00:53.180 --> 00:55.940
Which color is going to be applied to.
00:56.330 --> 00:59.210
And let's bring in our Lora adapters.
00:59.210 --> 00:59.790
There they are.
00:59.820 --> 01:06.090
You may remember that to be technically correct, there are in fact two adapters called A and B for
01:06.090 --> 01:10.380
each of the different, um target modules they have.
01:10.440 --> 01:15.420
Dimensionality given by R, and you may remember that alpha is the scaling factor.
01:15.420 --> 01:21.090
And it just is as simple as the way that these are applied to their target module is that it's alpha
01:21.120 --> 01:22.380
times A times B.
01:22.710 --> 01:24.450
Um simple as that.
01:24.450 --> 01:27.750
So that is our neural network that we know well.
01:27.780 --> 01:33.270
And we are going to be training the weights in this, these Laura adapters, which in our case is about
01:33.300 --> 01:36.090
109MB worth of, of weights.
01:36.780 --> 01:38.280
Uh, okay.
01:38.280 --> 01:40.290
So then what happens?
01:40.290 --> 01:42.900
We have an input prompt over on the left.
01:42.900 --> 01:47.760
It is like price is dollars and then it's the next token.
01:47.760 --> 01:50.070
We want the model to get good at predicting.
01:50.520 --> 01:57.780
The forward pass is when we take that and we in inference mode basically we we put it take it through
01:57.780 --> 02:02.400
the the model to say please predict the next token.
02:03.360 --> 02:09.780
And so that goes through the model and what comes out the other side is a predicted next token price
02:09.780 --> 02:10.260
is.
02:10.290 --> 02:11.820
And then 99.
02:11.820 --> 02:17.970
And again because we've got we're taking advantage of this simplicity of llama 3.1 that in fact 99 is
02:17.970 --> 02:19.170
just one token.
02:19.320 --> 02:20.400
Um, not not.
02:20.430 --> 02:23.730
And that will always be the case for any three digit number.
02:23.730 --> 02:27.300
That's not that that's critical, but it does simplify things a bit for us.
02:27.300 --> 02:29.190
So that's the forward pass.
02:29.220 --> 02:29.820
All right.
02:29.850 --> 02:32.040
Now onto the loss calculation.
02:32.610 --> 02:33.780
So here we are again.
02:33.780 --> 02:35.400
We've predicted the next token.
02:35.400 --> 02:41.400
So now the model is able to look up what or not the model the training process.
02:41.400 --> 02:44.010
The SFT trainer in our case looks up.
02:44.010 --> 02:45.960
What was the actual next token.
02:45.960 --> 02:49.080
Because we've got the training data, we know the actual next token.
02:49.080 --> 02:50.100
And what was it?
02:50.100 --> 02:51.810
Let's say it was 89.
02:51.810 --> 02:52.590
It was lower.
02:52.590 --> 02:54.000
So we were wrong.
02:54.000 --> 02:55.800
Wrong by $10.
02:56.070 --> 02:57.370
Um, or wrong.
02:57.370 --> 03:00.280
By different token, it doesn't know that this represents $10.
03:00.280 --> 03:02.350
It just knows it's a different token.
03:02.710 --> 03:05.650
Um, and so there is some kind of a loss.
03:05.650 --> 03:10.180
And in just a moment I'm going to explain what that loss is and why it's not quite as simple as it just
03:10.180 --> 03:11.500
being a different token.
03:11.500 --> 03:16.300
There's a technicality there that we'll get to, but for now you can just think of it as it predicted
03:16.300 --> 03:16.960
99.
03:16.990 --> 03:18.670
The actual value is 89.
03:18.700 --> 03:20.050
We have a loss.
03:20.680 --> 03:23.140
So that's the loss calculation.
03:23.140 --> 03:29.350
Step three is the backward pass that you hear people calling backprop or backward propagation.
03:29.350 --> 03:30.580
Back propagation.
03:30.760 --> 03:38.470
Um, and in backprop basically we look back through the network, back we go and we say, all right,
03:38.470 --> 03:43.930
so how much if we if we were to tweak these weights by a little bit, how much would that affect the
03:43.930 --> 03:44.650
loss?
03:44.650 --> 03:48.850
How sensitive is the loss to to those those weights.
03:48.880 --> 03:54.280
It gives us what we call the gradients of the weights of the parameters.
03:54.280 --> 03:57.320
So when I say weights are synonymous with parameters.
03:57.560 --> 04:00.770
Uh, so, uh, how, uh, um, yeah.
04:00.800 --> 04:02.120
What are the gradients?
04:02.120 --> 04:05.240
If we were to change those weights, what would it do to the loss?
04:05.240 --> 04:08.030
Because we want to try and improve things a little bit.
04:08.360 --> 04:14.120
Um, and so those, uh, red triangles are meant to represent Delta, uh, for showing like a gradient
04:14.210 --> 04:16.100
calculation that's happened there.
04:17.180 --> 04:22.730
Uh, and then finally we get to the last step, step four optimization.
04:22.940 --> 04:23.960
Here it comes.
04:23.960 --> 04:26.360
So we've got these gradients.
04:26.420 --> 04:30.770
And what we now need to do is we want to take a tiny step in the right direction.
04:30.770 --> 04:36.830
So we want to update the parameters in our Laura matrices a little tiny bit.
04:36.830 --> 04:41.720
So that next time if it gets the same input prompt, the loss will be a bit lower.
04:41.720 --> 04:42.890
It would do a bit better.
04:42.890 --> 04:48.200
So we're taking a step in the right direction, and we use the learning rate to decide how much of a
04:48.200 --> 04:48.980
step to take.
04:48.980 --> 04:53.810
Because as I said before, there are pros and cons of taking smaller or larger steps.
04:54.090 --> 05:00.690
Um, and you may remember that the optimizer we're using, the Adam W optimizer, does something quite
05:00.690 --> 05:06.720
cunning, where it doesn't just use these gradients, it keeps a kind of rolling average of prior gradients,
05:06.720 --> 05:12.540
so that it's being really smart about how to take that step in a way that's most likely to improve things.
05:12.660 --> 05:17.700
And it's also trying to make sure that we don't do things like overfitting and other dangers that we've
05:17.730 --> 05:18.810
talked about before.
05:19.200 --> 05:20.370
So we do that.
05:20.370 --> 05:26.430
And what happens is that the Laura adapters are then improved, and as they are applied in the future
05:26.430 --> 05:29.700
to our base model, it will do slightly better.
05:29.730 --> 05:36.390
The loss will be a bit lower next time, and that is why you see that the training loss is coming down
05:36.390 --> 05:40.740
and our wiggly charts, because it's learning and getting better and better.
05:41.070 --> 05:46.050
And so it's worth noting, of course, that the thing that the weights, the parameters that get changed
05:46.050 --> 05:51.180
are the parameters in the Laura adapters, the green ones, we don't actually change any parameters
05:51.180 --> 05:55.620
in the llama 3.1 base that is just too big, too many parameters.
05:55.620 --> 06:02.280
If you were doing ordinary training, fine tuning, not not Laura based, but ordinary, then it would
06:02.280 --> 06:09.120
be the whole llama 3.1 model that would need to be having gradients calculated and need to be shifted
06:09.120 --> 06:10.320
during optimization.
06:10.320 --> 06:13.650
And that, of course, is what these big companies like meta.
06:13.680 --> 06:19.170
That's how they've trained llama 3.1in the first place, and they've spent significant amounts of money
06:19.200 --> 06:25.620
doing it, which which we don't have easy access to, which is why we're using our Lora adapters instead.
06:26.190 --> 06:32.160
So that's the quick summary of optimization in the sorry of the whole of the training process.
06:32.160 --> 06:37.530
And I imagine that was probably mostly clear to you already, but I'm hoping that these diagrams of
06:37.530 --> 06:41.310
crystallized it for you and mean that everything has fallen into place.
06:41.310 --> 06:47.550
And in the next video, I'm just going to explain one more technicality, a very important technicality
06:47.550 --> 06:51.570
about what we mean by the prediction and the loss calculation.
06:51.600 --> 06:52.380
See you then.