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.
 
 

166 lines
4.9 KiB

WEBVTT
00:00.800 --> 00:01.310
Well.
00:01.310 --> 00:02.450
Hello there everybody.
00:02.450 --> 00:06.770
I am so grateful that you've made it through to the start of week seven.
00:06.800 --> 00:10.670
It is tremendous and I plan to not disappoint.
00:10.670 --> 00:12.530
I will make it worth your while.
00:12.530 --> 00:13.190
It's going.
00:13.220 --> 00:15.830
There's some really great content ahead.
00:15.830 --> 00:22.250
You may be feeling a bit of an anti-climax from what happened last week with our disappointing results.
00:22.250 --> 00:25.760
Fine tuning frontier model after so much hype.
00:25.850 --> 00:31.040
But what I can tell you is it's like a good movie where there's a period of time where things feel tense
00:31.040 --> 00:32.210
and disappointing.
00:32.420 --> 00:35.870
There's going to be redemption, there's going to be redemption.
00:35.870 --> 00:41.480
We will see if you're like me and you kept hammering away after after those results, you will have,
00:41.480 --> 00:45.260
I expect, got to a point where you have got some improvements.
00:45.260 --> 00:50.960
You get to a point where you can fine tune and actually get better results than we originally got with
00:51.200 --> 00:58.910
the Untuned version, but not significantly so, but a little bit better than I did in the last day.
00:58.910 --> 01:02.260
But anyways, for now we are moving swiftly onwards.
01:02.260 --> 01:08.860
We are launching into week seven and what a week do I have in store for you?
01:09.010 --> 01:10.780
Uh, it is a massive week.
01:10.780 --> 01:12.400
It is a really massive week.
01:12.400 --> 01:15.220
It is an advanced week where we're raising the bar.
01:15.220 --> 01:22.600
We're getting to a point where you're now acquiring advanced skills in this kind of deep learning field
01:22.600 --> 01:23.920
of of llms.
01:23.920 --> 01:29.860
And as we move more deeply into training, what you can already do, of course, you can already generate
01:29.890 --> 01:35.350
text and code with frontier models, with APIs, and with hugging face open source code.
01:35.350 --> 01:41.080
You can build rag pipelines and then you can now build data sets.
01:41.380 --> 01:47.260
Quite a long time we spent on data sets and baseline models and frontier models that you can fine tune,
01:47.260 --> 01:49.450
albeit with not amazing results.
01:49.450 --> 01:53.830
For our use case today we get advanced.
01:53.860 --> 01:59.160
We talk about using a technique called Lora for fine tuning open source models.
01:59.160 --> 02:00.870
We describe quantization.
02:00.870 --> 02:09.480
We talk about Q Laura, and we're getting into three key hyperparameters R alpha and target modules,
02:09.780 --> 02:13.980
which if you've not encountered them before, sounds like something straight out of Star Trek, but
02:13.980 --> 02:17.460
in fact will make complete sense by the end of today.
02:17.580 --> 02:23.550
And before we launch into it, I do need to take a moment to remind you of your eight week plan, and
02:23.580 --> 02:29.130
this may feel like this is boring for you at this point, but I think it's important to take pride in
02:29.130 --> 02:33.960
your accomplishments and to remember where everything that you've learned, you started.
02:34.290 --> 02:41.220
When we we first had our session to talk about, uh, uh, the, the opportunity with Llms and Frontier
02:41.250 --> 02:49.290
models and we compared them, um, we then looked at using the UIs, we use Gradio, we used agent ization
02:49.290 --> 02:50.460
Multi-modality.
02:50.490 --> 02:55.530
Then we got to hugging face the pipelines and then Tokenizers and models.
02:55.530 --> 03:01.900
Then in week four, we generated code by selecting LMS for the task.
03:01.900 --> 03:04.060
In week five, that was rag week.
03:04.060 --> 03:09.580
We used the wonderful Chroma and Lang chain to build some rag pipelines there, and hopefully you did
03:09.580 --> 03:12.580
the project, which I'm really excited to see what people do there.
03:12.820 --> 03:16.510
Uh, and then last week we fine tuned a frontier model.
03:16.510 --> 03:17.980
Uh, now you know how to do it.
03:17.980 --> 03:19.660
You understand about creating the data sets.
03:19.660 --> 03:23.950
You know how it works, and you understand when it's most applicable, when you're dealing with that
03:23.950 --> 03:28.870
kind of nuance of tone, and perhaps when it's not so applicable, in our case, when you're trying
03:28.870 --> 03:33.820
to actually build in functionality because a frontier model is already so massive.
03:34.270 --> 03:41.470
Um, so now we arrive at week seven, we are going to be fine tuning an open source model, one that
03:41.470 --> 03:46.540
is significantly smaller than a frontier model, and see where we get to with that.
03:46.540 --> 03:49.120
And week eight is where it all comes together.
03:49.210 --> 03:55.300
And so with that introduction, the main topic of the day is Laura, and we will launch right into that.