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.
 
 

97 lines
2.6 KiB

WEBVTT
00:01.400 --> 00:02.960
What more is there to say, really?
00:02.990 --> 00:04.820
Tomorrow is the day for results.
00:04.850 --> 00:07.610
A day that very excited indeed about.
00:07.850 --> 00:08.570
Uh.
00:08.570 --> 00:15.920
But I will mention everything that you can do at this point generating text and code with frontier models
00:15.920 --> 00:22.550
and their APIs, open source models through hugging faces, various libraries, tools, assistance,
00:22.580 --> 00:23.420
rag.
00:23.540 --> 00:25.460
It's all part of your skill set.
00:25.550 --> 00:32.420
You can follow the five step strategy I went through to solve problems including curating data, building
00:32.420 --> 00:39.320
a baseline model, and fine tuning frontier model with example training sets in those jsonl the JSON
00:39.320 --> 00:48.530
lines files that we uploaded, and now and now you can run Q Laura fine tuning for open source models,
00:48.560 --> 00:52.250
including defining and selecting your hyperparameters.
00:52.250 --> 00:56.270
And then you can run training and you can monitor training.
00:56.270 --> 01:05.900
And a great joy it is to uh, so tomorrow, you will be able to run inference on a fine tuned model,
01:05.900 --> 01:10.340
which isn't as simple as it sounds, because you can't just load it in.
01:10.370 --> 01:15.710
You have to know how to load in the base model and apply these chullora weights on top of it.
01:15.890 --> 01:16.730
But it's not that difficult.
01:16.730 --> 01:17.810
But but it is.
01:17.840 --> 01:19.400
There are a couple of steps to it.
01:19.850 --> 01:23.480
Uh, and then, of course, we'll take a moment to look at the results.
01:23.630 --> 01:31.220
Um, but most importantly, at that point, you will be able to carry out the end to end process from
01:31.220 --> 01:37.760
first idea and thought and definition of the commercial problem through to having a trained model,
01:37.850 --> 01:45.050
uh, to build your own proprietary verticalized LLM to solve a business problem.
01:45.650 --> 01:50.450
And I was going to try and put a third bullet in there, because I've normally had three bullets against
01:50.450 --> 01:52.100
the what you'll be able to do tomorrow.
01:52.100 --> 01:57.320
But it seems to me like that second bullet is enough of a mic drop that it's like you don't need a third
01:57.320 --> 01:57.740
bullet.
01:57.740 --> 01:58.820
That's all you need.
01:58.820 --> 02:00.830
That second bullet, it says it all.
02:00.830 --> 02:02.150
That's what you're going to be able to do.
02:02.150 --> 02:03.560
And it's huge.
02:03.590 --> 02:04.520
I'll see you then.