CMU Advanced NLP 2024 (7) Prompting/transcript.vtt

WEBVTT

00:00:01.319 --> 00:00:07.560
um today I want to talk about prompting

00:00:03.919 --> 00:00:09.639
and uh prompting is kind of a new uh

00:00:07.560 --> 00:00:11.320
Paradigm as of a few years ago with

00:00:09.639 --> 00:00:15.120
interacting with models it's now kind of

00:00:11.320 --> 00:00:16.880
the standard uh in doing so and

00:00:15.120 --> 00:00:19.880
basically what we do is we encourage a

00:00:16.880 --> 00:00:21.840
pre-trained model to make predictions by

00:00:19.880 --> 00:00:24.039
providing a textual prompt specifying

00:00:21.840 --> 00:00:25.960
the task to be done this is how you

00:00:24.039 --> 00:00:28.960
always interact with chat GPT or

00:00:25.960 --> 00:00:33.200
anything else like this

00:00:28.960 --> 00:00:36.200
um so prompting fundamentals uh the way

00:00:33.200 --> 00:00:38.360
that basic prompting works is you append

00:00:36.200 --> 00:00:42.079
a textual string to the beginning of the

00:00:38.360 --> 00:00:44.079
output and you complete it and exactly

00:00:42.079 --> 00:00:45.800
how you complete it can be based on any

00:00:44.079 --> 00:00:48.800
of the generation methods that we talked

00:00:45.800 --> 00:00:51.559
about in the previous class uh you know

00:00:48.800 --> 00:00:55.160
beam search it can be uh sampling it can

00:00:51.559 --> 00:00:58.480
be MBR or self-consistency or whatever

00:00:55.160 --> 00:01:00.960
else um so I I put in when a dog sees a

00:00:58.480 --> 00:01:03.680
squirrel it will usually

00:01:00.960 --> 00:01:06.280
um into gpt2 small which is a very small

00:01:03.680 --> 00:01:08.960
language model says Be Afraid of

00:01:06.280 --> 00:01:10.560
Anything unusual as an exception that's

00:01:08.960 --> 00:01:13.720
when a squirrel is usually afraid to

00:01:10.560 --> 00:01:16.280
bitee um so as you can see if the model

00:01:13.720 --> 00:01:19.560
is not super great you get a kind of not

00:01:16.280 --> 00:01:24.119
very great response also um but then I

00:01:19.560 --> 00:01:25.960
CED it into gp2 XL and uh what it says

00:01:24.119 --> 00:01:28.159
when a dog sees a squirrel it will

00:01:25.960 --> 00:01:30.640
usually lick the squirrel it will also

00:01:28.159 --> 00:01:34.000
touch its nose to the squirrel the tail

00:01:30.640 --> 00:01:37.880
and nose if it can um which might be

00:01:34.000 --> 00:01:40.280
true um one thing I I should note is

00:01:37.880 --> 00:01:43.040
when I generated these I used uh like

00:01:40.280 --> 00:01:45.200
actual regular ancestral sampling so I

00:01:43.040 --> 00:01:47.159
set the temperature to one I didn't do

00:01:45.200 --> 00:01:49.600
top feed didn't do top K or anything

00:01:47.159 --> 00:01:51.040
like this so this is a raw view of what

00:01:49.600 --> 00:01:53.799
the language model thinks is like

00:01:51.040 --> 00:01:58.479
actually a reasonable answer um if I

00:01:53.799 --> 00:02:00.159
modified the code to do something else

00:01:58.479 --> 00:02:02.560
actually maybe I can I can do that that

00:02:00.159 --> 00:02:04.960
right now but if I modified the code to

00:02:02.560 --> 00:02:08.879
use a

00:02:04.960 --> 00:02:12.119
different output we can actually see uh

00:02:08.879 --> 00:02:12.119
the different result that we

00:02:13.599 --> 00:02:17.959
get since I I have it here

00:02:18.360 --> 00:02:23.879
anyway actually sorry I'll need to

00:02:20.360 --> 00:02:27.239
modify the code on my my screen here

00:02:23.879 --> 00:02:32.120
um so I will

00:02:27.239 --> 00:02:35.040
set uh top K to 50 top P to

00:02:32.120 --> 00:02:38.360
0.95 so you see I I changed the

00:02:35.040 --> 00:02:38.360
generation parameters

00:02:38.760 --> 00:02:46.400
here and I'll uh run all of

00:02:43.159 --> 00:02:50.319
them you can see the uh the result that

00:02:46.400 --> 00:02:51.840
we get in a little bit but basically um

00:02:50.319 --> 00:02:54.800
so this is the standard method for

00:02:51.840 --> 00:02:57.319
prompting I intentionally use gpt2 small

00:02:54.800 --> 00:02:58.800
and gpt2 XL here because these are raw

00:02:57.319 --> 00:03:01.879
based language models they were just

00:02:58.800 --> 00:03:05.440
pre-trained as language models and so

00:03:01.879 --> 00:03:06.920
when we prompt them we're getting a

00:03:05.440 --> 00:03:09.200
language model that was just trained on

00:03:06.920 --> 00:03:12.280
lots of texts view of what is likely

00:03:09.200 --> 00:03:13.760
next text um there are other ways to

00:03:12.280 --> 00:03:15.599
train language models like instruction

00:03:13.760 --> 00:03:18.040
tuning and rlf which I'm going to be

00:03:15.599 --> 00:03:19.480
talking in future classes and if that's

00:03:18.040 --> 00:03:21.760
the case you might get a different

00:03:19.480 --> 00:03:23.159
response here so when a dog sees a

00:03:21.760 --> 00:03:25.720
squirrel it will usually get angry

00:03:23.159 --> 00:03:27.319
scratched the squirrel and run off uh

00:03:25.720 --> 00:03:29.080
some dogs may also attempt to capture

00:03:27.319 --> 00:03:30.799
the squirrel or attempt to eat it dogs

00:03:29.080 --> 00:03:32.599
will often to pick up the squirrel and

00:03:30.799 --> 00:03:36.400
eat it

00:03:32.599 --> 00:03:40.680
for it was more uh more violent than I

00:03:36.400 --> 00:03:44.280
expected any um

00:03:40.680 --> 00:03:45.720
so but anyway I think that like actually

00:03:44.280 --> 00:03:47.080
you can see that when I used the

00:03:45.720 --> 00:03:48.920
different generation parameters it

00:03:47.080 --> 00:03:51.480
actually gave me something that was

00:03:48.920 --> 00:03:54.319
maybe more typical than lick so lick is

00:03:51.480 --> 00:03:56.840
maybe a kind of unusual uh answer here

00:03:54.319 --> 00:03:58.680
but anyway

00:03:56.840 --> 00:04:03.040
cool

00:03:58.680 --> 00:04:05.680
so that's the basic idea of prompting we

00:04:03.040 --> 00:04:08.480
tend to use prompting to try to solve

00:04:05.680 --> 00:04:10.680
problems also so it's not just to

00:04:08.480 --> 00:04:14.200
complete text although completing text

00:04:10.680 --> 00:04:17.320
is useful and important like I complete

00:04:14.200 --> 00:04:19.199
text in my Gmail all the time uh you

00:04:17.320 --> 00:04:20.600
know it it's constantly giving me

00:04:19.199 --> 00:04:23.440
suggestions about what I should write

00:04:20.600 --> 00:04:24.800
next and I do tab autoc complete um you

00:04:23.440 --> 00:04:28.040
know on your phone you're doing that

00:04:24.800 --> 00:04:29.919
that's also using a language model um

00:04:28.040 --> 00:04:32.320
but very often we'll use prompting to do

00:04:29.919 --> 00:04:34.440
things other than just completing Texs

00:04:32.320 --> 00:04:36.000
and when we do this uh this is kind of

00:04:34.440 --> 00:04:38.199
the standard workflow for how we solve

00:04:36.000 --> 00:04:41.280
NLP tasks with prompting the way we do

00:04:38.199 --> 00:04:43.360
this is we fill in a prompt template

00:04:41.280 --> 00:04:46.080
predict the answer and post-process the

00:04:43.360 --> 00:04:46.080
answer in some

00:04:46.320 --> 00:04:51.880
way so prompt templates are templates

00:04:49.280 --> 00:04:55.280
where you will actually uh that you will

00:04:51.880 --> 00:04:57.479
fill in with an actual input and so if

00:04:55.280 --> 00:05:00.479
we have an input X which is something

00:04:57.479 --> 00:05:04.880
like I love this movie our template will

00:05:00.479 --> 00:05:08.360
be something like X overall it was Z or

00:05:04.880 --> 00:05:10.680
overall it was and so if we do that when

00:05:08.360 --> 00:05:13.320
we actually want to make a prediction we

00:05:10.680 --> 00:05:14.840
will uh convert this into the actual

00:05:13.320 --> 00:05:16.880
prompt we feed into the language model

00:05:14.840 --> 00:05:20.639
by filling in the template um I love

00:05:16.880 --> 00:05:24.919
this movie overall it was blank and then

00:05:20.639 --> 00:05:24.919
fill this uh continuation

00:05:25.840 --> 00:05:31.919
in a particular variety uh

00:05:30.000 --> 00:05:34.039
that we use very broadly nowadays

00:05:31.919 --> 00:05:36.240
because a lot of models are trained as

00:05:34.039 --> 00:05:38.240
chatbots um but actually even if they're

00:05:36.240 --> 00:05:41.199
not trained as chatbots this still works

00:05:38.240 --> 00:05:46.199
to some extent um is a chat

00:05:41.199 --> 00:05:49.919
prompt and so usually the way we we do

00:05:46.199 --> 00:05:53.240
this is we specify inputs in a format

00:05:49.919 --> 00:05:55.800
called the open AI messages format and

00:05:53.240 --> 00:05:58.199
uh this is this is what it looks like

00:05:55.800 --> 00:06:03.759
each we have a

00:05:58.199 --> 00:06:07.680
list of outputs each list is given a

00:06:03.759 --> 00:06:10.280
role and content and here so we have the

00:06:07.680 --> 00:06:12.479
role of system and the content is please

00:06:10.280 --> 00:06:15.319
classify movie reviews as positive or

00:06:12.479 --> 00:06:17.400
negative uh then we have the role user

00:06:15.319 --> 00:06:21.039
uh this movie is a

00:06:17.400 --> 00:06:24.919
banger um and then we have roles uh

00:06:21.039 --> 00:06:27.240
system message uh so is the roles we

00:06:24.919 --> 00:06:29.639
have the system and the system is a

00:06:27.240 --> 00:06:31.560
message provided to the system to

00:06:29.639 --> 00:06:33.560
influence Its Behavior it's to explain

00:06:31.560 --> 00:06:39.240
to it

00:06:33.560 --> 00:06:40.840
like how it should be working um and so

00:06:39.240 --> 00:06:43.199
you can see that this is explaining to

00:06:40.840 --> 00:06:46.400
the system how it should be working user

00:06:43.199 --> 00:06:48.680
is the message input by the user um and

00:06:46.400 --> 00:06:51.160
so this could be just a single message

00:06:48.680 --> 00:06:53.520
or if you have a multi-turn dialogue it

00:06:51.160 --> 00:06:55.080
can be like user and then assistant and

00:06:53.520 --> 00:06:56.680
then user and then assistant and then

00:06:55.080 --> 00:06:59.400
user and then assistant and that makes

00:06:56.680 --> 00:07:00.680
it clear that it's a multi-term dialogue

00:06:59.400 --> 00:07:02.800
so if you have a multi-term dialogue in

00:07:00.680 --> 00:07:06.319
chat GPT that's how they're feeding it

00:07:02.800 --> 00:07:06.319
in um into the

00:07:06.479 --> 00:07:12.440
system so what's happening behind the

00:07:08.840 --> 00:07:14.160
scenes with these chat prompts basically

00:07:12.440 --> 00:07:17.720
they're being converted into token

00:07:14.160 --> 00:07:19.680
strings and then fed into the model so

00:07:17.720 --> 00:07:21.800
despite the fact that this is fed in in

00:07:19.680 --> 00:07:23.560
this format and it makes you think that

00:07:21.800 --> 00:07:25.120
maybe something special is going on

00:07:23.560 --> 00:07:28.360
actually in most cases these are just

00:07:25.120 --> 00:07:30.199
being fed into the model uh as a prompt

00:07:28.360 --> 00:07:34.560
so these are just kind of special

00:07:30.199 --> 00:07:36.879
version of a uh of a template so here we

00:07:34.560 --> 00:07:40.560
have um this is what the Llama template

00:07:36.879 --> 00:07:43.319
looks like so basically you have um

00:07:40.560 --> 00:07:46.560
square bracket ins and then for the

00:07:43.319 --> 00:07:49.280
system message it's like um like angle

00:07:46.560 --> 00:07:51.240
bracket uh angle bracket sis uh close

00:07:49.280 --> 00:07:53.720
angle bracket close angle bracket and

00:07:51.240 --> 00:07:55.759
then the actual system message and then

00:07:53.720 --> 00:07:58.479
you have uh this closing out the system

00:07:55.759 --> 00:08:01.240
message this closing out the instruction

00:07:58.479 --> 00:08:04.120
then the user is surrounded by inst and

00:08:01.240 --> 00:08:06.599
then the assistant is just like a

00:08:04.120 --> 00:08:08.400
regular string so this is what the

00:08:06.599 --> 00:08:12.319
actual textual string that's fed into

00:08:08.400 --> 00:08:14.199
llama chat models is we can contrast

00:08:12.319 --> 00:08:19.440
that to some other models so alpaka

00:08:14.199 --> 00:08:22.400
looks like this um uh so we have like

00:08:19.440 --> 00:08:24.879
hash instruction colon and then the

00:08:22.400 --> 00:08:26.639
instruction for the user there there's

00:08:24.879 --> 00:08:28.879
no distinction between system and user

00:08:26.639 --> 00:08:31.960
so it's like hash instruction and then

00:08:28.879 --> 00:08:35.240
the user message and then hash response

00:08:31.960 --> 00:08:37.760
and then be assistant so it's not super

00:08:35.240 --> 00:08:39.640
important which one we use here um the

00:08:37.760 --> 00:08:41.919
important thing is that this matches

00:08:39.640 --> 00:08:44.039
with what uh the model is trained and

00:08:41.919 --> 00:08:46.640
I'll show you some example uh you know

00:08:44.039 --> 00:08:50.680
I'll talk about that in more detail

00:08:46.640 --> 00:08:52.880
later and we have a reference uh that I

00:08:50.680 --> 00:08:56.600
got this uh

00:08:52.880 --> 00:08:58.519
from and there's this toolkit that I um

00:08:56.600 --> 00:09:02.680
I rather like recently it's called light

00:08:58.519 --> 00:09:05.079
llm it makes it very easy to uh query

00:09:02.680 --> 00:09:07.240
different llms uh and kind of like

00:09:05.079 --> 00:09:09.320
unified things so basically you can

00:09:07.240 --> 00:09:11.800
query many different types of LMS like

00:09:09.320 --> 00:09:14.440
open AI or open source models or other

00:09:11.800 --> 00:09:17.079
things like that and what happens behind

00:09:14.440 --> 00:09:19.120
the scene is it basically takes um the

00:09:17.079 --> 00:09:20.839
open AI messages format and converts it

00:09:19.120 --> 00:09:22.880
into the appropriate prompt format for

00:09:20.839 --> 00:09:24.680
whatever model you're using or the

00:09:22.880 --> 00:09:27.120
appropriate API calls for whatever thing

00:09:24.680 --> 00:09:29.800
you're using but

00:09:27.120 --> 00:09:31.399
um this here basically

00:09:29.800 --> 00:09:33.800
um if you click through this link shows

00:09:31.399 --> 00:09:35.959
you okay this is what it looks like for

00:09:33.800 --> 00:09:37.880
alpaca um so you have the instruction

00:09:35.959 --> 00:09:40.920
instruction response this is what it

00:09:37.880 --> 00:09:44.880
looks like for llama 2 chat this is what

00:09:40.920 --> 00:09:48.480
it looks like for the oama um for AMA

00:09:44.880 --> 00:09:49.920
this is what it looks like for mistol

00:09:48.480 --> 00:09:52.160
and other things like that so you see

00:09:49.920 --> 00:09:53.440
all of these are very similar but

00:09:52.160 --> 00:09:55.000
they're like slightly different and

00:09:53.440 --> 00:09:58.120
getting these right is actually kind of

00:09:55.000 --> 00:10:01.120
important for the model doing a good

00:09:58.120 --> 00:10:01.120
job

00:10:03.640 --> 00:10:10.399
um any questions about

00:10:05.880 --> 00:10:15.360
this yeah like say you start PR with

00:10:10.399 --> 00:10:18.160
this um inut and then you started simar

00:10:15.360 --> 00:10:21.320
without

00:10:18.160 --> 00:10:24.640
model could you give an example yeah so

00:10:21.320 --> 00:10:28.040
say um my account is a great movie or

00:10:24.640 --> 00:10:31.040
this movie is great in front of I put

00:10:28.040 --> 00:10:31.040
UMR

00:10:34.279 --> 00:10:39.519
model

00:10:36.399 --> 00:10:42.440
so depend it depends a lot on the

00:10:39.519 --> 00:10:45.959
bottle the reason why this system

00:10:42.440 --> 00:10:48.720
message was input here in the first

00:10:45.959 --> 00:10:52.440
place was this wasn't originally a

00:10:48.720 --> 00:10:54.240
feature of open AI models uh open AI was

00:10:52.440 --> 00:10:56.440
the first place to introduce this which

00:10:54.240 --> 00:10:58.519
is why I I'm calling it open ey messages

00:10:56.440 --> 00:10:59.800
formul they didn't originally have

00:10:58.519 --> 00:11:02.360
something like this but they were having

00:10:59.800 --> 00:11:04.360
lots of trouble with um people trying to

00:11:02.360 --> 00:11:07.600
reveal the prompts that were given to

00:11:04.360 --> 00:11:09.680
systems uh like called like prompt

00:11:07.600 --> 00:11:12.040
injection attacks or like jailbreaking

00:11:09.680 --> 00:11:15.399
attacks or stff like that and so the

00:11:12.040 --> 00:11:17.079
models would basically reveal this

00:11:15.399 --> 00:11:19.600
prompt that was being used behind the

00:11:17.079 --> 00:11:22.760
scenes by whatever customer of open a

00:11:19.600 --> 00:11:26.120
was like deploying a system and so in

00:11:22.760 --> 00:11:29.120
order to fix this basically what open AI

00:11:26.120 --> 00:11:30.480
did I believe I believe like they're

00:11:29.120 --> 00:11:32.279
don't actually tell you exactly what

00:11:30.480 --> 00:11:36.040
they did ever but I'm assuming what they

00:11:32.279 --> 00:11:37.680
did is they trained uh their models so

00:11:36.040 --> 00:11:39.240
that the models would not output

00:11:37.680 --> 00:11:41.639
anything that's included in the system

00:11:39.240 --> 00:11:43.839
message so the system message is used to

00:11:41.639 --> 00:11:46.120
influence behavior but it like they're

00:11:43.839 --> 00:11:48.200
explicitly trained to not output things

00:11:46.120 --> 00:11:49.880
that are included in there and so if you

00:11:48.200 --> 00:11:53.360
put the

00:11:49.880 --> 00:11:56.200
actual if you put the actual thing that

00:11:53.360 --> 00:11:59.639
you wanted to evaluate within the system

00:11:56.200 --> 00:12:01.839
message it might still predict

00:11:59.639 --> 00:12:04.839
the sentiment correctly but it won't

00:12:01.839 --> 00:12:06.920
repeat the the stuff that was in system

00:12:04.839 --> 00:12:09.920
message

00:12:06.920 --> 00:12:09.920
B

00:12:14.160 --> 00:12:20.480
yeah after we give it the yeah yeah so

00:12:18.320 --> 00:12:23.040
the that's a great question so typically

00:12:20.480 --> 00:12:26.480
this is hand created so you you create

00:12:23.040 --> 00:12:29.680
something like this um I I have a a

00:12:26.480 --> 00:12:32.120
bracket X here but another way people

00:12:29.680 --> 00:12:33.800
typically specify this is you just have

00:12:32.120 --> 00:12:36.880
a

00:12:33.800 --> 00:12:41.199
big um you just have a big python string

00:12:36.880 --> 00:12:41.199
which is like um you know

00:12:42.040 --> 00:12:46.480
please um please

00:12:49.279 --> 00:12:55.440
specify and then you

00:12:52.440 --> 00:12:55.440
have

00:12:56.160 --> 00:13:02.240
um and then you substitute in uh like

00:12:59.880 --> 00:13:04.440
the input into this place here so you

00:13:02.240 --> 00:13:07.760
usually handw write it I'm going to

00:13:04.440 --> 00:13:07.760
talk excuse

00:13:07.800 --> 00:13:14.120
me and to end about some methods to

00:13:10.320 --> 00:13:16.120
learn these also but um I'd say like 90

00:13:14.120 --> 00:13:18.320
95% of the time people are just writing

00:13:16.120 --> 00:13:18.320
the

00:13:19.959 --> 00:13:24.560
man yep I would

00:13:25.920 --> 00:13:31.639
write

00:13:27.760 --> 00:13:31.639
and real input that

00:13:33.240 --> 00:13:38.040
I yeah so typically the template is

00:13:36.360 --> 00:13:39.800
written when you decide what system you

00:13:38.040 --> 00:13:41.839
want to create so you decide you want to

00:13:39.800 --> 00:13:44.519
create a sentiment analysis system so

00:13:41.839 --> 00:13:46.760
you create a template that either says

00:13:44.519 --> 00:13:48.079
like please classify the topic in the

00:13:46.760 --> 00:13:50.959
case of a model that was trained to

00:13:48.079 --> 00:13:52.240
follow instructions or if you have a

00:13:50.959 --> 00:13:54.240
base model that was not trained to

00:13:52.240 --> 00:13:58.079
follow instructions which is rare rare

00:13:54.240 --> 00:14:00.279
nowadays but gpd2 or La llama 2 without

00:13:58.079 --> 00:14:02.320
chat tuning is as an example of that

00:14:00.279 --> 00:14:05.600
then you would need to create a template

00:14:02.320 --> 00:14:10.040
that looks like this um where

00:14:05.600 --> 00:14:11.360
you put the model in a situation where

00:14:10.040 --> 00:14:13.839
the

00:14:11.360 --> 00:14:15.240
next word that follows up should be

00:14:13.839 --> 00:14:17.120
indicative of the answer to your

00:14:15.240 --> 00:14:20.120
question so like positive or negative or

00:14:17.120 --> 00:14:21.800
something like that so um but either way

00:14:20.120 --> 00:14:24.639
like usually you handw write this when

00:14:21.800 --> 00:14:27.199
you decide what task is you want to do

00:14:24.639 --> 00:14:29.000
then this input X this comes at test

00:14:27.199 --> 00:14:32.920
time this comes when you actually Dey

00:14:29.000 --> 00:14:34.240
your system um so this would be like an

00:14:32.920 --> 00:14:37.040
Amazon review that you wanted to

00:14:34.240 --> 00:14:37.040
classify using an

00:14:37.720 --> 00:14:42.720
image cool any other

00:14:40.519 --> 00:14:46.480
questions okay let's

00:14:42.720 --> 00:14:48.160
move um so basically this is what is

00:14:46.480 --> 00:14:49.920
happening behind the scenes I don't know

00:14:48.160 --> 00:14:53.040
what open AI format is because they

00:14:49.920 --> 00:14:54.639
won't tell us of course um but you know

00:14:53.040 --> 00:14:56.000
I'm assuming that that's similar to

00:14:54.639 --> 00:14:59.399
what's happening in

00:14:56.000 --> 00:15:01.959
op okay um so the next thing that we do

00:14:59.399 --> 00:15:05.360
is answer prediction so given uh The

00:15:01.959 --> 00:15:08.320
Prompt we predict the answer um and so

00:15:05.360 --> 00:15:11.880
using whatever algorithm we want to use

00:15:08.320 --> 00:15:14.880
uh we predict you know fantastic

00:15:11.880 --> 00:15:14.880
here

00:15:15.120 --> 00:15:21.639
um and actually it might not predict

00:15:19.959 --> 00:15:26.399
fantastic it might predict something

00:15:21.639 --> 00:15:28.120
else like overall it was um a really

00:15:26.399 --> 00:15:30.000
fantastic movie that I liked a lot or

00:15:28.120 --> 00:15:33.839
something like so it might also do

00:15:30.000 --> 00:15:36.880
something like that so based on that we

00:15:33.839 --> 00:15:39.600
want to select the actual output out of

00:15:36.880 --> 00:15:41.160
the generated uh outputs and I'm calling

00:15:39.600 --> 00:15:43.639
this uh

00:15:41.160 --> 00:15:45.959
postprocessing so for instance we might

00:15:43.639 --> 00:15:48.240
take the output as is so for something

00:15:45.959 --> 00:15:50.880
like just you interacting with chat

00:15:48.240 --> 00:15:53.360
jpt um or interacting with a chat model

00:15:50.880 --> 00:15:55.639
you might be looking at the text as is

00:15:53.360 --> 00:15:58.319
or it might be formatting the output for

00:15:55.639 --> 00:16:00.079
easy Vis visualization selecting only

00:15:58.319 --> 00:16:02.440
parts of the output that you want to use

00:16:00.079 --> 00:16:04.560
or mapping the output to other

00:16:02.440 --> 00:16:07.600
actions so to give an example of

00:16:04.560 --> 00:16:10.079
formatting this is a feature of uh chat

00:16:07.600 --> 00:16:13.440
GPT or Bard or any that you interact

00:16:10.079 --> 00:16:14.920
with but um I wrote please write a table

00:16:13.440 --> 00:16:18.759
with the last five presidents and their

00:16:14.920 --> 00:16:20.319
birth dates and chat GPT is happy to do

00:16:18.759 --> 00:16:22.000
this for me it says here is a table with

00:16:20.319 --> 00:16:24.920
the last five US presidents and their

00:16:22.000 --> 00:16:27.639
birth dates um Joe Biden Donald Trump

00:16:24.920 --> 00:16:31.720
Barack Obama George W wish Bill Clinton

00:16:27.639 --> 00:16:33.600
um but this is written in markdown um or

00:16:31.720 --> 00:16:35.079
I assume it's written in markdown so it

00:16:33.600 --> 00:16:37.880
basically makes this table and then

00:16:35.079 --> 00:16:39.319
renders it in an easy to view way so

00:16:37.880 --> 00:16:41.000
this is really important if you're

00:16:39.319 --> 00:16:42.440
building a user facing system because

00:16:41.000 --> 00:16:44.279
you want to be able to render these

00:16:42.440 --> 00:16:46.279
things but the only thing a large

00:16:44.279 --> 00:16:48.880
language model can output is text right

00:16:46.279 --> 00:16:50.279
it can output a string of tokens so uh

00:16:48.880 --> 00:16:54.000
this is a really good way to interact

00:16:50.279 --> 00:16:55.759
with it um I I followed by saying output

00:16:54.000 --> 00:16:58.720
that in Json format so it says here's

00:16:55.759 --> 00:17:00.360
the information in Json format and

00:16:58.720 --> 00:17:02.000
instead of just giving me a big Json

00:17:00.360 --> 00:17:04.199
string it gives me syntax highlighting

00:17:02.000 --> 00:17:06.880
and all the other stuff like this um

00:17:04.199 --> 00:17:09.760
presumably what it's doing here is it's

00:17:06.880 --> 00:17:12.839
outputting um like a triple hash or

00:17:09.760 --> 00:17:15.160
something like this um the reason why I

00:17:12.839 --> 00:17:17.600
know that is because

00:17:15.160 --> 00:17:21.079
like seems to be making a mistake down

00:17:17.600 --> 00:17:23.280
here for some reason um like uh

00:17:21.079 --> 00:17:25.079
outputting a weird Le formatted thing at

00:17:23.280 --> 00:17:26.160
that and so even chat GPT makes mistakes

00:17:25.079 --> 00:17:30.320
some of the

00:17:26.160 --> 00:17:32.400
time um

00:17:30.320 --> 00:17:33.960
cool um another thing that you might

00:17:32.400 --> 00:17:35.520
want to do is especially if you're not

00:17:33.960 --> 00:17:37.360
using it in like a a directly

00:17:35.520 --> 00:17:40.200
user-facing application but you want to

00:17:37.360 --> 00:17:41.840
use it to extract some information or

00:17:40.200 --> 00:17:45.440
make some classification decision or

00:17:41.840 --> 00:17:47.280
something like that um you often select

00:17:45.440 --> 00:17:49.880
information that's indicative of the

00:17:47.280 --> 00:17:52.360
answer and so I love this movie overall

00:17:49.880 --> 00:17:53.960
it was a movie that was simply fantastic

00:17:52.360 --> 00:17:56.600
um you can do things like extract

00:17:53.960 --> 00:17:59.440
keywords like fantastic and use that to

00:17:56.600 --> 00:18:01.360
indicate positive sentiment

00:17:59.440 --> 00:18:04.080
there's various methods for doing this

00:18:01.360 --> 00:18:05.919
and these are also used in the

00:18:04.080 --> 00:18:08.679
benchmarks that are used to evaluate

00:18:05.919 --> 00:18:09.799
language models so it's you know like

00:18:08.679 --> 00:18:11.039
even if you're not building an

00:18:09.799 --> 00:18:12.679
application directly but you're just

00:18:11.039 --> 00:18:14.120
trying to do well in this class and get

00:18:12.679 --> 00:18:15.679
like a high score on a leaderboard or

00:18:14.120 --> 00:18:20.320
something it's still useful to know

00:18:15.679 --> 00:18:22.159
about these things so um for things like

00:18:20.320 --> 00:18:24.039
classification um you can identify

00:18:22.159 --> 00:18:27.159
keywords like fantastic that might be

00:18:24.039 --> 00:18:29.120
indicative of the class another thing

00:18:27.159 --> 00:18:31.559
that's uh pretty common is for

00:18:29.120 --> 00:18:34.480
regression or numerical problems you

00:18:31.559 --> 00:18:37.440
identify numbers and pull out the

00:18:34.480 --> 00:18:40.400
numbers and use those numbers as the

00:18:37.440 --> 00:18:42.360
answer um for code uh you can pull out

00:18:40.400 --> 00:18:45.080
code Snippets and triple back ticks and

00:18:42.360 --> 00:18:46.960
then execute the code for example so all

00:18:45.080 --> 00:18:48.600
of these things are basically heuristic

00:18:46.960 --> 00:18:50.159
methods but they can be used to pull out

00:18:48.600 --> 00:18:53.440
the actual answer that you want from the

00:18:50.159 --> 00:18:53.440
text that's generated di

00:18:54.480 --> 00:19:00.320
know cool uh any questions about that

00:19:02.280 --> 00:19:07.880
the final thing is output mapping um

00:19:04.640 --> 00:19:11.120
given an answer uh map it into a class

00:19:07.880 --> 00:19:13.360
label or a continuous value and so this

00:19:11.120 --> 00:19:16.000
is doing something like taking fantastic

00:19:13.360 --> 00:19:18.480
and mapping it into the class

00:19:16.000 --> 00:19:21.000
positive uh and so you know if we want

00:19:18.480 --> 00:19:23.000
to extract fi one to five star ratings

00:19:21.000 --> 00:19:25.559
from reviews this is something you would

00:19:23.000 --> 00:19:29.360
need to do and very often it's like a

00:19:25.559 --> 00:19:33.880
one to um one class to

00:19:29.360 --> 00:19:35.720
many um many word mapping and uh by

00:19:33.880 --> 00:19:37.400
doing this you can basically get a more

00:19:35.720 --> 00:19:38.720
robust mapping onto the number that you

00:19:37.400 --> 00:19:42.400
actually

00:19:38.720 --> 00:19:42.400
want I actually

00:19:42.720 --> 00:19:48.919
coincidentally on uh on Twitter saw a

00:19:45.280 --> 00:19:48.919
really good example of this like a week

00:19:55.880 --> 00:20:00.520
ago and yeah I don't know if I'm going

00:19:59.120 --> 00:20:05.440
to be able to find it in a reasonable

00:20:00.520 --> 00:20:08.520
time frame but basically um there was

00:20:05.440 --> 00:20:11.080
a person who was using gp4 to create a

00:20:08.520 --> 00:20:14.120
model uh to like reward open source

00:20:11.080 --> 00:20:15.880
models for good and bad you know

00:20:14.120 --> 00:20:18.320
responses

00:20:15.880 --> 00:20:20.799
and they started out with giving it a

00:20:18.320 --> 00:20:24.480
one to five star rating and then they

00:20:20.799 --> 00:20:28.360
switched it into very good good okay bad

00:20:24.480 --> 00:20:31.280
very bad and then um then asked to

00:20:28.360 --> 00:20:34.520
generate you know those like very good

00:20:31.280 --> 00:20:37.039
good bad okay bad very bad instead of

00:20:34.520 --> 00:20:40.360
one to five and that worked a lot better

00:20:37.039 --> 00:20:43.480
like the GPT model was a lot more uh

00:20:40.360 --> 00:20:46.039
like likely to get the answer correct um

00:20:43.480 --> 00:20:48.880
than it was if you gave a one to five

00:20:46.039 --> 00:20:50.799
star rating so this is something you

00:20:48.880 --> 00:20:54.280
should think about pretty seriously and

00:20:50.799 --> 00:20:57.440
the way you can think about it is How

00:20:54.280 --> 00:20:59.679
likely was this data to appear in a

00:20:57.440 --> 00:21:02.520
large Corp of data on the

00:20:59.679 --> 00:21:04.760
internet and it might be like a lot less

00:21:02.520 --> 00:21:08.679
likely that it's like how good is this

00:21:04.760 --> 00:21:11.400
movie five then how good is this movie

00:21:08.679 --> 00:21:13.960
really good like just think of like the

00:21:11.400 --> 00:21:16.200
occurrence probability and you can even

00:21:13.960 --> 00:21:18.600
um like mine this data from the the web

00:21:16.200 --> 00:21:21.320
if you want to to try to find out the

00:21:18.600 --> 00:21:24.520
best you know

00:21:21.320 --> 00:21:30.039
like the best things

00:21:24.520 --> 00:21:30.039
there cool um any questions about this

00:21:35.360 --> 00:21:39.480
yeah how is

00:21:37.720 --> 00:21:43.039
it

00:21:39.480 --> 00:21:45.919
learning so the model the model is

00:21:43.039 --> 00:21:47.600
predicting txt and like accurately it's

00:21:45.919 --> 00:21:50.200
not even predicting the word fantastic

00:21:47.600 --> 00:21:54.480
it's predicting the token ID like

00:21:50.200 --> 00:21:57.600
73521 or something like that um but you

00:21:54.480 --> 00:21:58.679
know if it has seen that token ID more

00:21:57.600 --> 00:22:00.840
frequent

00:21:58.679 --> 00:22:04.240
after reviews than it has seen the token

00:22:00.840 --> 00:22:06.000
ID for the number one or the number five

00:22:04.240 --> 00:22:07.520
then it's more likely to predict that

00:22:06.000 --> 00:22:10.279
accurately right it's more likely to

00:22:07.520 --> 00:22:11.880
predict fantastic than it is to predict

00:22:10.279 --> 00:22:14.679
five star or something like that just

00:22:11.880 --> 00:22:16.720
because fantastic is more frequent and

00:22:14.679 --> 00:22:18.880
so because of that if you think about

00:22:16.720 --> 00:22:22.120
like what has it seen in all of the data

00:22:18.880 --> 00:22:24.240
on the internet and like model your um

00:22:22.120 --> 00:22:26.960
model your answers here appropriately

00:22:24.240 --> 00:22:28.520
then that can give you

00:22:26.960 --> 00:22:30.320
betters

00:22:28.520 --> 00:22:32.120
this is a very important rule of thumb

00:22:30.320 --> 00:22:33.400
like don't try to make a language model

00:22:32.120 --> 00:22:35.039
do something it's never seen in the

00:22:33.400 --> 00:22:38.200
pre-training data and it will make your

00:22:35.039 --> 00:22:40.240
life a lot easier so um you can think

00:22:38.200 --> 00:22:41.880
that going forward

00:22:40.240 --> 00:22:44.679
to

00:22:41.880 --> 00:22:48.559
cool so next I want to move into fat

00:22:44.679 --> 00:22:49.679
prompting or in context learning um so

00:22:48.559 --> 00:22:52.159
fat

00:22:49.679 --> 00:22:54.440
prompting basically what we do is we

00:22:52.159 --> 00:22:55.799
provide a few examples of the task

00:22:54.440 --> 00:22:58.440
together with the

00:22:55.799 --> 00:23:00.080
instruction and the way this work works

00:22:58.440 --> 00:23:02.360
is you write an instruction like please

00:23:00.080 --> 00:23:05.919
classify movie reviews as positive or

00:23:02.360 --> 00:23:08.120
negative and add like input uh I really

00:23:05.919 --> 00:23:10.320
don't like this movie output negative uh

00:23:08.120 --> 00:23:12.480
input this movie is great output

00:23:10.320 --> 00:23:16.640
positive

00:23:12.480 --> 00:23:18.880
and this is um pretty effective the

00:23:16.640 --> 00:23:21.799
thing it's most effective for are

00:23:18.880 --> 00:23:24.400
twofold it's most effective for making

00:23:21.799 --> 00:23:26.360
sure that you get the formatting right

00:23:24.400 --> 00:23:27.640
uh because if you have a few examples

00:23:26.360 --> 00:23:28.679
the model will tend to follow those

00:23:27.640 --> 00:23:30.840
examples

00:23:28.679 --> 00:23:34.440
with respect to formatting especially if

00:23:30.840 --> 00:23:37.320
we're talking about like gp4 models um

00:23:34.440 --> 00:23:40.400
or strong GPT models it's also effective

00:23:37.320 --> 00:23:42.400
if you're using weaker models so like

00:23:40.400 --> 00:23:44.720
stronger models like gp4 tend to be

00:23:42.400 --> 00:23:46.720
pretty good at following instructions so

00:23:44.720 --> 00:23:49.520
if you say

00:23:46.720 --> 00:23:51.640
um please classify movie reviews as

00:23:49.520 --> 00:23:54.000
positive or negative it will be more

00:23:51.640 --> 00:23:56.279
likely to just output positive or

00:23:54.000 --> 00:23:58.760
negative um but if you have weaker

00:23:56.279 --> 00:24:01.720
models it might say I really don't like

00:23:58.760 --> 00:24:03.559
this movie output uh I think I think

00:24:01.720 --> 00:24:05.640
this is probably negative or something

00:24:03.559 --> 00:24:07.240
like that it will you know it might not

00:24:05.640 --> 00:24:10.080
follow the instructions as well and it's

00:24:07.240 --> 00:24:14.240
more effective to provide as in context

00:24:10.080 --> 00:24:17.600
examples um so so this is a one uh

00:24:14.240 --> 00:24:19.480
one uh thing to remember one thing I

00:24:17.600 --> 00:24:22.120
should mention also is when I say F shot

00:24:19.480 --> 00:24:25.720
prompting and in context learning these

00:24:22.120 --> 00:24:27.880
are basically the same thing uh they

00:24:25.720 --> 00:24:29.720
basically refer to the same concept but

00:24:27.880 --> 00:24:31.919
just from slightly different

00:24:29.720 --> 00:24:34.799
examples uh from sorry slightly

00:24:31.919 --> 00:24:36.919
different angles PE shot is in contrast

00:24:34.799 --> 00:24:39.320
to zero shot so zero shot means you're

00:24:36.919 --> 00:24:43.039
providing no examples so zero shot

00:24:39.320 --> 00:24:45.720
prompting you would have none uh few

00:24:43.039 --> 00:24:47.240
shot you have several examples in

00:24:45.720 --> 00:24:49.679
context learning means that you're

00:24:47.240 --> 00:24:51.640
learning how to do a task but instead of

00:24:49.679 --> 00:24:54.320
providing the model with fine-tuning

00:24:51.640 --> 00:24:56.679
data you're providing the examples in

00:24:54.320 --> 00:24:58.080
the language models context so they both

00:24:56.679 --> 00:25:00.919
basically mean the same thing but

00:24:58.080 --> 00:25:03.159
they're they're just contrasting to like

00:25:00.919 --> 00:25:06.559
either a zero shot or fine tuning which

00:25:03.159 --> 00:25:06.559
is why the terminology is

00:25:06.880 --> 00:25:13.520
different so they usering interface

00:25:11.320 --> 00:25:16.080
and for the

00:25:13.520 --> 00:25:17.760
rendering uh yes you can definitely do F

00:25:16.080 --> 00:25:20.039
shot prompting I'm actually going to

00:25:17.760 --> 00:25:23.440
talk exactly about exactly how you do

00:25:20.039 --> 00:25:26.320
this in like an open AI model um here

00:25:23.440 --> 00:25:28.240
which is for open AI models there's a

00:25:26.320 --> 00:25:31.320
couple ways that you could do this one

00:25:28.240 --> 00:25:33.640
way you could do this is you could um

00:25:31.320 --> 00:25:36.279
you could have the role be user and the

00:25:33.640 --> 00:25:39.279
role be assistant and just add like

00:25:36.279 --> 00:25:41.159
additional conversational history into

00:25:39.279 --> 00:25:43.159
the the messages that you're sending to

00:25:41.159 --> 00:25:46.240
the language model but actually the

00:25:43.159 --> 00:25:49.120
recommended way of doing this um which

00:25:46.240 --> 00:25:51.880
is in the openi cookbook uh which is in

00:25:49.120 --> 00:25:53.919
the reference is that you send this as a

00:25:51.880 --> 00:25:58.200
system message but you provide this like

00:25:53.919 --> 00:26:00.840
additional name variable here um with

00:25:58.200 --> 00:26:02.840
example user and example assistant the

00:26:00.840 --> 00:26:06.200
main reason why you do this is just

00:26:02.840 --> 00:26:08.080
because if you don't um if you send it

00:26:06.200 --> 00:26:10.600
in as the like user and assistant the

00:26:08.080 --> 00:26:12.799
model might refer back to the few shot

00:26:10.600 --> 00:26:14.320
examples as something that happened

00:26:12.799 --> 00:26:15.760
previously in the conversation whereas

00:26:14.320 --> 00:26:18.200
if you send it in the system message

00:26:15.760 --> 00:26:19.799
it's guaranteed to not do that so I

00:26:18.200 --> 00:26:23.600
think it's like less of an accuracy

00:26:19.799 --> 00:26:26.360
thing it's more of a like it's more of a

00:26:23.600 --> 00:26:29.120
privacy prompt privacy thing uh than

00:26:26.360 --> 00:26:30.880
anything else so this is a recommended

00:26:29.120 --> 00:26:33.159
way of doing this on the other hand if

00:26:30.880 --> 00:26:34.600
you're using like an open source model

00:26:33.159 --> 00:26:36.600
uh you need to be careful because this

00:26:34.600 --> 00:26:38.279
name might not even be included in the

00:26:36.600 --> 00:26:40.080
prompt template like for example in the

00:26:38.279 --> 00:26:41.840
light llm prompt templates that I was

00:26:40.080 --> 00:26:44.080
sending in this is not even included at

00:26:41.840 --> 00:26:46.480
all so you might just get a weird system

00:26:44.080 --> 00:26:49.720
message that uh is poorly fored so you

00:26:46.480 --> 00:26:53.600
need to be a little bit conscious

00:26:49.720 --> 00:26:55.799
this um cool any questions here does

00:26:53.600 --> 00:26:58.880
that answer the

00:26:55.799 --> 00:27:02.279
question okay

00:26:58.880 --> 00:27:05.000
um so one one thing to be aware of is

00:27:02.279 --> 00:27:07.039
llms are sensitive to small changes and

00:27:05.000 --> 00:27:12.080
in context examples that you provide to

00:27:07.039 --> 00:27:14.600
them so uh previous work has examined

00:27:12.080 --> 00:27:19.399
this from a number of angles there's a

00:27:14.600 --> 00:27:22.679
paper by Luol and they examine the

00:27:19.399 --> 00:27:25.000
sensitivity to example ordering so like

00:27:22.679 --> 00:27:28.399
if you take the same examples and you

00:27:25.000 --> 00:27:30.840
just order them in different orders um

00:27:28.399 --> 00:27:32.679
you can actually get very wildly

00:27:30.840 --> 00:27:35.600
different

00:27:32.679 --> 00:27:37.520
results um and this is especially true

00:27:35.600 --> 00:27:40.320
for smaller models so the smaller models

00:27:37.520 --> 00:27:42.720
here are like the gpt2 models the larger

00:27:40.320 --> 00:27:47.440
models here are like the GPT the larger

00:27:42.720 --> 00:27:47.440
model here is GPT 3.5 uh I

00:27:48.399 --> 00:27:54.120
believe other things that people have

00:27:50.559 --> 00:27:56.760
looked at are label balance so um how

00:27:54.120 --> 00:27:58.559
important is it for the labels to be

00:27:56.760 --> 00:28:01.440
balanced

00:27:58.559 --> 00:28:02.799
um and if you're doing sentiment

00:28:01.440 --> 00:28:05.240
classification for example you might

00:28:02.799 --> 00:28:07.519
have only positive examples or only

00:28:05.240 --> 00:28:10.000
negative examples and if you have only

00:28:07.519 --> 00:28:13.279
positive or negative examples this can

00:28:10.000 --> 00:28:15.559
uh help or hurt your accuracy uh for

00:28:13.279 --> 00:28:17.200
example on this Amazon review data set

00:28:15.559 --> 00:28:18.679
most of the reviews are positive so you

00:28:17.200 --> 00:28:20.840
actually do better by having lots of

00:28:18.679 --> 00:28:23.640
positive examples in your in context

00:28:20.840 --> 00:28:26.600
examples on the other hand for sst2 this

00:28:23.640 --> 00:28:29.159
is label balanced so having only

00:28:26.600 --> 00:28:31.799
positive or negative is worse on average

00:28:29.159 --> 00:28:34.279
than having three positive and one

00:28:31.799 --> 00:28:36.679
negative another thing is label coverage

00:28:34.279 --> 00:28:38.679
so if we're talking about multi class

00:28:36.679 --> 00:28:41.120
classification um

00:28:38.679 --> 00:28:42.919
having good coverage of all of the

00:28:41.120 --> 00:28:45.919
classes that you want to include in your

00:28:42.919 --> 00:28:49.120
multiclass classification is important

00:28:45.919 --> 00:28:51.720
um to some extent but if you have uh

00:28:49.120 --> 00:28:53.440
more uh you can also confuse some model

00:28:51.720 --> 00:28:55.840
especially if they're minority labels so

00:28:53.440 --> 00:28:57.799
if you have a whole bunch of like random

00:28:55.840 --> 00:28:59.080
minority labels and that can cause so

00:28:57.799 --> 00:29:01.399
this is something important to think

00:28:59.080 --> 00:29:04.640
about if you're planning on solving kind

00:29:01.399 --> 00:29:08.640
of like classification tests um I I've

00:29:04.640 --> 00:29:11.000
also had my own experience with uh using

00:29:08.640 --> 00:29:13.159
GPT for evaluation for machine

00:29:11.000 --> 00:29:14.760
translation and when we use GPT for

00:29:13.159 --> 00:29:18.559
evaluation for machine translation it

00:29:14.760 --> 00:29:20.799
was very important to add um like high

00:29:18.559 --> 00:29:22.760
uh high scoring values low score high

00:29:20.799 --> 00:29:26.320
scoring outputs low scoring outputs some

00:29:22.760 --> 00:29:27.840
in the middle um and so it's also the

00:29:26.320 --> 00:29:30.760
case for regression

00:29:27.840 --> 00:29:30.760
uh problems as

00:29:32.600 --> 00:29:37.320
well cool um any questions

00:29:38.159 --> 00:29:45.000
here um however this is not super

00:29:42.240 --> 00:29:46.600
predictable um so there's not like any

00:29:45.000 --> 00:29:48.399
rule of thumb that tells you like this

00:29:46.600 --> 00:29:49.720
is or as far as I know there's not any

00:29:48.399 --> 00:29:51.640
rule of thumb that tells you this is the

00:29:49.720 --> 00:29:54.000
way you should construct in context

00:29:51.640 --> 00:29:55.880
examples uh there are lots of papers

00:29:54.000 --> 00:29:57.799
that say they have methods that work

00:29:55.880 --> 00:30:01.000
better but I don't know if there's any

00:29:57.799 --> 00:30:02.559
like gold standard IND indry practice

00:30:01.000 --> 00:30:05.799
for doing something like this at the

00:30:02.559 --> 00:30:07.799
moment so just to give an example uh

00:30:05.799 --> 00:30:10.399
this paper it's a really nice paper

00:30:07.799 --> 00:30:13.440
examining why uh in context Learning

00:30:10.399 --> 00:30:17.279
Works one thing one interesting finding

00:30:13.440 --> 00:30:19.760
that they have is they output they take

00:30:17.279 --> 00:30:22.720
in context examples but they randomize

00:30:19.760 --> 00:30:27.320
the labels they make the labels wrong

00:30:22.720 --> 00:30:29.519
some of the time so even with completely

00:30:27.320 --> 00:30:32.120
wrong labels even with labels that are

00:30:29.519 --> 00:30:34.399
correct 0% of the time you still get

00:30:32.120 --> 00:30:37.360
much much better accuracy than if you

00:30:34.399 --> 00:30:39.440
use no Inc context examples and why is

00:30:37.360 --> 00:30:41.640
this probably you know it's getting the

00:30:39.440 --> 00:30:44.600
model formatting correct it's getting

00:30:41.640 --> 00:30:47.679
like the names of the labels correct

00:30:44.600 --> 00:30:49.039
even if it's not uh accurate so it seems

00:30:47.679 --> 00:30:50.519
like it's not really using these for

00:30:49.039 --> 00:30:52.640
training data it's using them more just

00:30:50.519 --> 00:30:56.240
to know the formatting

00:30:52.640 --> 00:30:59.399
appropriate like

00:30:56.240 --> 00:31:01.399
you so you already

00:30:59.399 --> 00:31:03.760
have

00:31:01.399 --> 00:31:08.840
right how is it

00:31:03.760 --> 00:31:11.240
Ma like is it just y one y i gu I'm just

00:31:08.840 --> 00:31:15.000
ask how you would inter

00:31:11.240 --> 00:31:16.480
that so this is you're not training the

00:31:15.000 --> 00:31:17.880
model at the moment we're going to talk

00:31:16.480 --> 00:31:19.360
about that next class but right now

00:31:17.880 --> 00:31:21.279
you're taking a model that has already

00:31:19.360 --> 00:31:22.840
been trained and you're providing it

00:31:21.279 --> 00:31:25.519
with a few examples and then you're

00:31:22.840 --> 00:31:28.679
asking it to fill in um the following

00:31:25.519 --> 00:31:30.880
examples just examples

00:31:28.679 --> 00:31:32.960
yes

00:31:30.880 --> 00:31:34.679
exactly and it's pretty amazing that

00:31:32.960 --> 00:31:36.440
that works in the first place especially

00:31:34.679 --> 00:31:39.840
with a model that hasn't been explicitly

00:31:36.440 --> 00:31:41.200
trained that way but um there's a a fair

00:31:39.840 --> 00:31:42.320
amount of research that I think we're

00:31:41.200 --> 00:31:43.960
probably going to be talking about in

00:31:42.320 --> 00:31:47.000
the interpretability class about why

00:31:43.960 --> 00:31:49.600
this happens but um

00:31:47.000 --> 00:31:51.279
basically my my interpretation for why

00:31:49.600 --> 00:31:53.679
this happens is because there's so much

00:31:51.279 --> 00:31:56.000
repetitive stuff on the internet right

00:31:53.679 --> 00:31:58.240
there's a bunch of examples of math

00:31:56.000 --> 00:32:00.399
problems which is like

00:31:58.240 --> 00:32:02.279
question one and then the math problem

00:32:00.399 --> 00:32:04.320
and then the answer question two math

00:32:02.279 --> 00:32:06.440
problem and then the answer so in order

00:32:04.320 --> 00:32:08.320
to model the text on the internet it

00:32:06.440 --> 00:32:12.120
needs to learn how to be able to do

00:32:08.320 --> 00:32:15.399
these things but so um

00:32:12.120 --> 00:32:17.760
cool the second thing is uh more

00:32:15.399 --> 00:32:20.000
demonstrations can sometimes hurt

00:32:17.760 --> 00:32:22.120
accuracy so this is like binary

00:32:20.000 --> 00:32:25.080
classification versus multiple choice

00:32:22.120 --> 00:32:27.440
question answering um and actually with

00:32:25.080 --> 00:32:30.919
binary classification the model ends up

00:32:27.440 --> 00:32:33.159
getting worse um with uh more examples

00:32:30.919 --> 00:32:36.799
probably just because the longer context

00:32:33.159 --> 00:32:39.320
uh you know confuses the model or moves

00:32:36.799 --> 00:32:41.320
the instructions that are provided to

00:32:39.320 --> 00:32:44.279
the model farther away in the context so

00:32:41.320 --> 00:32:48.120
it starts forgetting them so

00:32:44.279 --> 00:32:50.240
um basically what I want to say is uh

00:32:48.120 --> 00:32:51.760
you know this is more of an art than a

00:32:50.240 --> 00:32:53.279
science you might not get entirely

00:32:51.760 --> 00:32:55.840
predictable results but don't worry it's

00:32:53.279 --> 00:32:59.320
not just

00:32:55.840 --> 00:32:59.320
you cool cool

00:33:09.200 --> 00:33:15.320
yeah it can't so the question is if the

00:33:12.639 --> 00:33:17.039
in context examples reflect the data

00:33:15.320 --> 00:33:18.919
distribution well would that boost the

00:33:17.039 --> 00:33:24.240
accuracy I think the answer is probably

00:33:18.919 --> 00:33:26.039
yes yeah um I don't know if that it's

00:33:24.240 --> 00:33:27.679
that clear because like what I would

00:33:26.039 --> 00:33:29.919
expect

00:33:27.679 --> 00:33:33.559
is better

00:33:29.919 --> 00:33:37.240
coverage is probably more

00:33:33.559 --> 00:33:39.760
important than better representativeness

00:33:37.240 --> 00:33:41.960
so like even if you have some minority

00:33:39.760 --> 00:33:43.639
labels um it's probably better for the

00:33:41.960 --> 00:33:44.880
model to know what those minority labels

00:33:43.639 --> 00:33:47.279
look like and that's going to be

00:33:44.880 --> 00:33:49.120
especially true for like stronger models

00:33:47.279 --> 00:33:50.679
um I think

00:33:49.120 --> 00:33:54.320
so

00:33:50.679 --> 00:33:56.440
cool okay so uh next I want to talk

00:33:54.320 --> 00:33:59.000
about Chain of Thought prompting um so

00:33:56.440 --> 00:34:01.320
Chain of Thought prompting is a very

00:33:59.000 --> 00:34:04.080
popular way of prompting

00:34:01.320 --> 00:34:06.080
models and the way it works is you get

00:34:04.080 --> 00:34:07.839
the model to explain its reasoning

00:34:06.080 --> 00:34:12.679
before making an

00:34:07.839 --> 00:34:14.520
answer um and so sorry this example is a

00:34:12.679 --> 00:34:18.879
little bit small but like the standard

00:34:14.520 --> 00:34:20.480
prompting method is uh like Roger has

00:34:18.879 --> 00:34:22.000
five tennis balls he buys two more cans

00:34:20.480 --> 00:34:23.480
of tennis balls each can has three

00:34:22.000 --> 00:34:28.200
tennis balls how many tennis balls does

00:34:23.480 --> 00:34:29.359
he have now um the answer is 11 and so

00:34:28.200 --> 00:34:32.119
um this

00:34:29.359 --> 00:34:34.320
is an in context example and then you

00:34:32.119 --> 00:34:37.240
have your input which has a different

00:34:34.320 --> 00:34:39.000
problem uh the cafeteria has 23 apples

00:34:37.240 --> 00:34:40.639
if they Ed 20 to make lunch and bought

00:34:39.000 --> 00:34:41.800
six more how many apples do they have

00:34:40.639 --> 00:34:46.720
the answer is

00:34:41.800 --> 00:34:49.000
27 um and so this is wrong so what Chain

00:34:46.720 --> 00:34:52.000
of Thought prompting does is instead of

00:34:49.000 --> 00:34:54.960
just giving the answer it gives you an

00:34:52.000 --> 00:34:57.079
additional reasoning chain uh that says

00:34:54.960 --> 00:34:59.680
R started with five balls two cans of of

00:34:57.079 --> 00:35:01.800
three tennis balls uh each of six tennis

00:34:59.680 --> 00:35:04.520
balls 5 plus 6 equals 11 the answer is

00:35:01.800 --> 00:35:06.280
11 and so then when you feed this in

00:35:04.520 --> 00:35:08.000
basically the model will generate a

00:35:06.280 --> 00:35:10.240
similar reasoning chain and then it's

00:35:08.000 --> 00:35:13.400
more likely to get the answer correct

00:35:10.240 --> 00:35:15.720
and this very robustly works

00:35:13.400 --> 00:35:19.440
for many

00:35:15.720 --> 00:35:21.440
different problems where a reasoning

00:35:19.440 --> 00:35:23.520
chain is

00:35:21.440 --> 00:35:27.839
necessary and if you think about the

00:35:23.520 --> 00:35:30.359
reason why this uh why this works I

00:35:27.839 --> 00:35:33.040
think there's basically two reasons why

00:35:30.359 --> 00:35:34.440
um the first reason is I I only wrote

00:35:33.040 --> 00:35:36.560
one on the thing here but the first

00:35:34.440 --> 00:35:38.760
reason is it allows the model to

00:35:36.560 --> 00:35:41.359
decompose harder problems into simpler

00:35:38.760 --> 00:35:45.119
problems and simpler problems are easier

00:35:41.359 --> 00:35:47.560
right so um instead

00:35:45.119 --> 00:35:51.319
of immediately trying to solve the whole

00:35:47.560 --> 00:35:53.800
problem in a single go it will first

00:35:51.319 --> 00:35:56.520
solve the problem of like what how many

00:35:53.800 --> 00:35:58.920
are left after you use buy and so it

00:35:56.520 --> 00:36:00.240
gets three and so now it has this three

00:35:58.920 --> 00:36:02.480
here so now it can solve the next

00:36:00.240 --> 00:36:05.160
problem of adding six that's equal to 9

00:36:02.480 --> 00:36:07.880
so it's solving simpler sub problems

00:36:05.160 --> 00:36:11.440
than it is and uh compared to harder

00:36:07.880 --> 00:36:13.920
ones another reason why is it allows for

00:36:11.440 --> 00:36:17.319
adaptive computation time so if you

00:36:13.920 --> 00:36:17.319
think about like a Transformer

00:36:19.000 --> 00:36:23.119
model um if you think about a

00:36:21.280 --> 00:36:25.560
Transformer model a Transformer model

00:36:23.119 --> 00:36:27.200
has fixed computation time for

00:36:25.560 --> 00:36:29.920
predicting each token right a fixed

00:36:27.200 --> 00:36:31.560
number of layers it um and based on that

00:36:29.920 --> 00:36:33.839
fixed number of layers it passes all the

00:36:31.560 --> 00:36:36.520
information through and makes a

00:36:33.839 --> 00:36:38.200
prediction and some problems are harder

00:36:36.520 --> 00:36:39.599
than others right so it would be very

00:36:38.200 --> 00:36:42.480
wasteful to have a really big

00:36:39.599 --> 00:36:45.640
Transformer that could solve you know

00:36:42.480 --> 00:36:49.119
really complex math problems in the same

00:36:45.640 --> 00:36:53.359
amount of time it takes to predict that

00:36:49.119 --> 00:36:55.280
the next word is like uh dog after the

00:36:53.359 --> 00:36:57.280
word the big or something like that

00:36:55.280 --> 00:36:58.560
right so there are some things that are

00:36:57.280 --> 00:37:00.000
easy we can do in a second there are

00:36:58.560 --> 00:37:01.839
some things that take us more time and

00:37:00.000 --> 00:37:05.880
essentially this Chain of Thought

00:37:01.839 --> 00:37:09.280
reasoning is um is doing that it's

00:37:05.880 --> 00:37:12.280
giving it more time to solve the harder

00:37:09.280 --> 00:37:12.280
problems

00:37:17.200 --> 00:37:22.440
yes

00:37:18.839 --> 00:37:23.960
okay yeah good good question so so

00:37:22.440 --> 00:37:26.200
that's what um that's what this next

00:37:23.960 --> 00:37:27.920
paper does so uh the the question was

00:37:26.200 --> 00:37:31.160
what what happens if we just ask it to

00:37:27.920 --> 00:37:34.800
reason and the answer is it still works

00:37:31.160 --> 00:37:37.000
um and this paper was really like I I I

00:37:34.800 --> 00:37:39.760
love this paper for its Simplicity and

00:37:37.000 --> 00:37:43.160
cleverness and basically uh they

00:37:39.760 --> 00:37:45.000
contrast few shot learning few shot

00:37:43.160 --> 00:37:49.200
Chain of Thought where you provide Chain

00:37:45.000 --> 00:37:52.160
of Thought examples zero shot prompting

00:37:49.200 --> 00:37:54.560
basically and zero shot Chain of Thought

00:37:52.160 --> 00:37:58.720
So what they do is they just

00:37:54.560 --> 00:38:00.280
add uh let's thinks step by step that

00:37:58.720 --> 00:38:04.200
they add that phrase to the end of The

00:38:00.280 --> 00:38:06.079
Prompt and then that elicits the model

00:38:04.200 --> 00:38:08.000
to basically do Chain of Thought

00:38:06.079 --> 00:38:09.240
reasoning without any further examples

00:38:08.000 --> 00:38:12.599
of how that Chain of Thought reasoning

00:38:09.240 --> 00:38:14.440
works why does this work again because

00:38:12.599 --> 00:38:16.760
like on the internet there's a bunch of

00:38:14.440 --> 00:38:20.240
examples of math problem solving data

00:38:16.760 --> 00:38:22.800
sets or QA corpora where it says let

00:38:20.240 --> 00:38:24.480
things step by step and after that you

00:38:22.800 --> 00:38:28.040
you know consistently have this sort of

00:38:24.480 --> 00:38:29.800
resoning chain added there so um good

00:38:28.040 --> 00:38:31.200
good intuition uh that this paper

00:38:29.800 --> 00:38:32.480
answers the question and this like

00:38:31.200 --> 00:38:36.119
actually does

00:38:32.480 --> 00:38:39.119
work one interesting thing is

00:38:36.119 --> 00:38:39.119
um

00:38:39.720 --> 00:38:45.200
now if I go to chat

00:38:47.319 --> 00:38:52.240
GPT and I say

00:38:50.480 --> 00:38:58.520
um

00:38:52.240 --> 00:38:58.520
I am teaching a class with 98

00:38:58.720 --> 00:39:06.000
students

00:39:01.480 --> 00:39:08.400
70% turn in the

00:39:06.000 --> 00:39:10.720
assignment hint on

00:39:08.400 --> 00:39:17.720
time uh

00:39:10.720 --> 00:39:21.880
10% and it in play how many did not ENT

00:39:17.720 --> 00:39:21.880
in let's see let's see if this

00:39:25.079 --> 00:39:29.200
works okay it's writing code for

00:39:29.440 --> 00:39:34.599
me which is that's a feature slide I

00:39:32.119 --> 00:39:37.319
kind of didn't uh I kind of didn't

00:39:34.599 --> 00:39:37.319
wanted to do

00:39:40.040 --> 00:39:44.160
that okay

00:39:45.920 --> 00:39:50.359
um I do not

00:39:54.280 --> 00:39:59.720
like okay so um

00:39:57.040 --> 00:39:59.720
it's a little bit

00:40:15.720 --> 00:40:21.839
slow okay so there there that worked but

00:40:19.680 --> 00:40:24.640
not that I did not say let's think step

00:40:21.839 --> 00:40:27.760
by step I didn't I didn't ask it to do

00:40:24.640 --> 00:40:29.240
this um and the reason why is um we're

00:40:27.760 --> 00:40:31.119
going to talk about instruction tuning

00:40:29.240 --> 00:40:34.359
next time but basically GPT has been

00:40:31.119 --> 00:40:36.560
tuned to do this reasoning even if you

00:40:34.359 --> 00:40:38.480
don't ask it to do that uh it wouldn't

00:40:36.560 --> 00:40:40.839
do that naturally but it's because lots

00:40:38.480 --> 00:40:43.880
of supervised data has been added into

00:40:40.839 --> 00:40:46.920
this model so like another thing is like

00:40:43.880 --> 00:40:48.960
if you are planning on doing anything

00:40:46.920 --> 00:40:51.240
about like Chain of Thought reasoning or

00:40:48.960 --> 00:40:53.000
or stuff like that as a class project

00:40:51.240 --> 00:40:54.960
you need to keep in mind that the like

00:40:53.000 --> 00:40:58.280
GPD models have already been trained to

00:40:54.960 --> 00:41:00.040
do this and so so if you want to like

00:40:58.280 --> 00:41:01.599
try to find out a better way to elicit

00:41:00.040 --> 00:41:03.960
this from a raw model you'll need to use

00:41:01.599 --> 00:41:07.119
a raw model like llama 2 with no chat

00:41:03.960 --> 00:41:10.200
tuning or stuff like that um in order to

00:41:07.119 --> 00:41:12.520
uh do that in a neutral in a neutral

00:41:10.200 --> 00:41:14.960
setting that hasn't been contaminated by

00:41:12.520 --> 00:41:14.960
like super

00:41:15.960 --> 00:41:20.520
L cool um any

00:41:21.079 --> 00:41:27.720
questions okay um so next I want to talk

00:41:24.720 --> 00:41:31.280
about prompting in programs that uh Chad

00:41:27.720 --> 00:41:35.560
GPD gave me a good example of uh why why

00:41:31.280 --> 00:41:37.160
this is useful or important um so

00:41:35.560 --> 00:41:40.640
there's two results actually both of

00:41:37.160 --> 00:41:43.440
these are are from my uh collaborators

00:41:40.640 --> 00:41:45.839
but the first one is um it demonstrates

00:41:43.440 --> 00:41:48.720
that structuring outputs at programs can

00:41:45.839 --> 00:41:51.599
help you get better results even if the

00:41:48.720 --> 00:41:55.119
task isn't a programmatic task so this

00:41:51.599 --> 00:41:57.000
is kind of interesting um so we were

00:41:55.119 --> 00:41:59.319
looking at predicting stru structured

00:41:57.000 --> 00:42:01.640
outputs and these structured outputs

00:41:59.319 --> 00:42:03.839
specifically are procedural knowledge

00:42:01.640 --> 00:42:06.920
like this so like how do we cook a pie

00:42:03.839 --> 00:42:09.040
or how do we serve pot pies on a plate

00:42:06.920 --> 00:42:10.800
and we had this procedural knowledge

00:42:09.040 --> 00:42:14.040
like take the pies out to pool open the

00:42:10.800 --> 00:42:16.079
cabinet drawer take out several plates

00:42:14.040 --> 00:42:17.720
and we wanted to know the dependencies

00:42:16.079 --> 00:42:19.520
between these so we could create a

00:42:17.720 --> 00:42:22.559
structured like procedural knowledge

00:42:19.520 --> 00:42:25.599
base so this is not an inherently code

00:42:22.559 --> 00:42:27.200
based task it's not a you know you could

00:42:25.599 --> 00:42:28.880
just ask

00:42:27.200 --> 00:42:32.160
the model in natural language and that

00:42:28.880 --> 00:42:35.000
would work as well so we structured

00:42:32.160 --> 00:42:37.720
things in a couple varieties so we had a

00:42:35.000 --> 00:42:39.800
textual format we had uh something in

00:42:37.720 --> 00:42:43.079
the dot format which is a way to draw

00:42:39.800 --> 00:42:45.480
graphs and then we had we also tried

00:42:43.079 --> 00:42:47.240
structuring the output in Python so

00:42:45.480 --> 00:42:48.960
these are just different ways to format

00:42:47.240 --> 00:42:50.720
the output they all say the same thing

00:42:48.960 --> 00:42:54.599
and we can extract the answer from all

00:42:50.720 --> 00:42:56.920
of them um but we found that structuring

00:42:54.599 --> 00:42:58.480
it in in Python basically is the more

00:42:56.920 --> 00:43:02.920
effective way of doing

00:42:58.480 --> 00:43:04.680
this so why why is it this the case the

00:43:02.920 --> 00:43:06.280
answer is essentially the same thing

00:43:04.680 --> 00:43:08.680
that I was talking about before with you

00:43:06.280 --> 00:43:11.480
know predicting excellent instead of

00:43:08.680 --> 00:43:13.319
five right you know it's seen a ton of

00:43:11.480 --> 00:43:15.960
python in it's training data so it's

00:43:13.319 --> 00:43:17.760
very good at predicting python uh it's

00:43:15.960 --> 00:43:20.359
less good at predicting dot format

00:43:17.760 --> 00:43:24.240
because it seemed less do format and it

00:43:20.359 --> 00:43:26.640
hasn't seen very much text here

00:43:24.240 --> 00:43:29.960
um another

00:43:26.640 --> 00:43:32.359
comment is code is very highly

00:43:29.960 --> 00:43:33.559
structured compared to natural language

00:43:32.359 --> 00:43:35.599
and because code is very highly

00:43:33.559 --> 00:43:37.520
structured we have things like

00:43:35.599 --> 00:43:39.079
dependencies where we refer back to

00:43:37.520 --> 00:43:41.079
variables that we defined before and

00:43:39.079 --> 00:43:44.119
other things like this so I think when

00:43:41.079 --> 00:43:46.760
it starts outputting code the models get

00:43:44.119 --> 00:43:48.359
into this mode which say yes please

00:43:46.760 --> 00:43:51.280
refer back to the things you've seen

00:43:48.359 --> 00:43:53.440
previously more often like attend to

00:43:51.280 --> 00:43:57.040
previous stuff more often and don't just

00:43:53.440 --> 00:43:59.760
like generate things you know uh

00:43:57.040 --> 00:44:02.440
arbitrarily and hallucinate you know new

00:43:59.760 --> 00:44:04.119
content and because of this for

00:44:02.440 --> 00:44:05.559
generating structured outputs even if

00:44:04.119 --> 00:44:08.920
the structured outputs don't need to be

00:44:05.559 --> 00:44:11.520
code you can benefit by doing

00:44:08.920 --> 00:44:13.200
this another thing that's a really handy

00:44:11.520 --> 00:44:16.319
trick is anytime you want to get a

00:44:13.200 --> 00:44:19.079
structured output out of a model

00:44:16.319 --> 00:44:22.760
um you can ask it to generate something

00:44:19.079 --> 00:44:24.839
in Json instead of generating it in uh

00:44:22.760 --> 00:44:26.640
in text and the reason why Json is

00:44:24.839 --> 00:44:28.079
useful is you can press the on you can

00:44:26.640 --> 00:44:32.319
pull out the strings and other stuff

00:44:28.079 --> 00:44:34.839
like that um so this can be very

00:44:32.319 --> 00:44:37.960
effective because if you just add an

00:44:34.839 --> 00:44:40.200
instruction that says please um please

00:44:37.960 --> 00:44:42.440
format things in this particular

00:44:40.200 --> 00:44:43.680
format often the model won't listen to

00:44:42.440 --> 00:44:44.800
you and it will output something in a

00:44:43.680 --> 00:44:46.280
different format you need to write a

00:44:44.800 --> 00:44:48.599
really annoying parser to pull out the

00:44:46.280 --> 00:44:50.280
information that you actually want but

00:44:48.599 --> 00:44:51.960
it gets Json right almost all of the

00:44:50.280 --> 00:44:54.040
time just because it's seen so much Json

00:44:51.960 --> 00:44:57.880
so that's a nice trick if you want to do

00:44:54.040 --> 00:45:01.520
something like that

00:44:57.880 --> 00:45:03.559
another uh thing is a paper uh called

00:45:01.520 --> 00:45:08.079
program AED language models that we did

00:45:03.559 --> 00:45:10.200
about a year ago and the method that we

00:45:08.079 --> 00:45:13.760
proposed here is using a program to

00:45:10.200 --> 00:45:16.480
generate outputs uh using a program to

00:45:13.760 --> 00:45:19.440
generate outputs and this can be more

00:45:16.480 --> 00:45:22.319
precise than asking an LM to do so and

00:45:19.440 --> 00:45:26.720
so instead of doing Chain of Thought

00:45:22.319 --> 00:45:30.640
prompting we created a few F shot

00:45:26.720 --> 00:45:34.319
examples where we wrote like the text

00:45:30.640 --> 00:45:37.160
here and then the text in English and

00:45:34.319 --> 00:45:40.160
then we had code corresponding code the

00:45:37.160 --> 00:45:42.280
text in English corresponding code and

00:45:40.160 --> 00:45:44.960
then the answer is and then the final

00:45:42.280 --> 00:45:48.160
code and then we basically generate this

00:45:44.960 --> 00:45:49.640
code and execute it to get the answer so

00:45:48.160 --> 00:45:52.319
like as you saw this is implemented in

00:45:49.640 --> 00:45:54.280
chat GP now it's uh you write something

00:45:52.319 --> 00:45:56.319
out it will decide whether it wants to

00:45:54.280 --> 00:45:58.599
generate code or generate text depending

00:45:56.319 --> 00:46:00.760
on the type of problem and it's just

00:45:58.599 --> 00:46:03.559
more precise it can solve like actually

00:46:00.760 --> 00:46:05.200
rather complex problems like uh you know

00:46:03.559 --> 00:46:07.880
calculating how much tax you need to be

00:46:05.200 --> 00:46:10.880
paying or something like that

00:46:07.880 --> 00:46:12.480
um it's especially useful for numeric

00:46:10.880 --> 00:46:14.599
questions and it's implemented in things

00:46:12.480 --> 00:46:17.040
like the chat GPT uh code interpreter

00:46:14.599 --> 00:46:18.640
bar to execution other things like that

00:46:17.040 --> 00:46:22.079
it's pretty cool it can actually do

00:46:18.640 --> 00:46:24.440
visualizations for you for papers also

00:46:22.079 --> 00:46:28.000
so if you ask it to visualize data for

00:46:24.440 --> 00:46:30.200
you um chat GPD now does a pretty good

00:46:28.000 --> 00:46:32.640
job of doing this like to give an

00:46:30.200 --> 00:46:34.319
example I asked it I gave it a big

00:46:32.640 --> 00:46:35.760
python list and asked it to generate a

00:46:34.319 --> 00:46:37.839
histogram and it did a really good job

00:46:35.760 --> 00:46:40.240
of it for me it also gives you the code

00:46:37.839 --> 00:46:42.839
so you can go in and modify it later so

00:46:40.240 --> 00:46:44.720
um I would definitely recommend you know

00:46:42.839 --> 00:46:46.200
thinking about using this uh either in

00:46:44.720 --> 00:46:49.200
your research or just to write your

00:46:46.200 --> 00:46:51.480
reports uh for this class so um this

00:46:49.200 --> 00:46:55.839
class is uh generative AI friendly

00:46:51.480 --> 00:46:57.760
mostly so like I do I do want you to

00:46:55.839 --> 00:46:59.880
learn the things we expect you to learn

00:46:57.760 --> 00:47:02.480
which is why I suggest that you don't

00:46:59.880 --> 00:47:04.400
like just write every uh everything for

00:47:02.480 --> 00:47:06.280
assignment number one with chat GP key

00:47:04.400 --> 00:47:07.720
but I think even if you tried to do that

00:47:06.280 --> 00:47:09.640
it'd probably get it wrong in subtle

00:47:07.720 --> 00:47:10.920
ways so you're probably better off

00:47:09.640 --> 00:47:13.880
understanding the content

00:47:10.920 --> 00:47:16.400
anyway um

00:47:13.880 --> 00:47:18.160
cool this can also be expanded a whole

00:47:16.400 --> 00:47:21.559
lot into like agents and tools and I'm

00:47:18.160 --> 00:47:21.559
going to talk about that separately

00:47:22.800 --> 00:47:27.720
later cool uh any any things about this

00:47:29.040 --> 00:47:34.200
okay I'm uh I'm going to go next so

00:47:31.800 --> 00:47:36.079
prompt engineering um when you're

00:47:34.200 --> 00:47:37.280
designing prompts uh there's a number of

00:47:36.079 --> 00:47:38.240
different ways you can do this you can

00:47:37.280 --> 00:47:41.559
do this

00:47:38.240 --> 00:47:42.960
manually uh you to do this you configure

00:47:41.559 --> 00:47:44.520
a manual template based on the

00:47:42.960 --> 00:47:46.160
characteristics of the task using all of

00:47:44.520 --> 00:47:48.880
the knowledge that I told you

00:47:46.160 --> 00:47:50.119
before you can also do automated search

00:47:48.880 --> 00:47:52.079
and there's a number of different ways

00:47:50.119 --> 00:47:55.119
to do automated search for

00:47:52.079 --> 00:47:58.319
prompts uh the first one is doing some

00:47:55.119 --> 00:48:00.599
sort of search discret space uh so you

00:47:58.319 --> 00:48:02.720
find a prompt that is

00:48:00.599 --> 00:48:04.680
essentially

00:48:02.720 --> 00:48:06.640
text the other one is search in

00:48:04.680 --> 00:48:08.559
continuous space so you find a prompt

00:48:06.640 --> 00:48:10.680
that isn't actually comprehensible text

00:48:08.559 --> 00:48:14.760
but nonetheless is a good

00:48:10.680 --> 00:48:16.960
prompt so looking at manual engineering

00:48:14.760 --> 00:48:19.000
um making sure that the format matches

00:48:16.960 --> 00:48:21.680
that of a trained model uh such as the

00:48:19.000 --> 00:48:24.359
chat format is actually really really

00:48:21.680 --> 00:48:26.119
important um and this can have a a large

00:48:24.359 --> 00:48:28.119
effect on models there's a really paper

00:48:26.119 --> 00:48:30.000
that demonstrated this convincingly

00:48:28.119 --> 00:48:33.200
before and also releases some software

00:48:30.000 --> 00:48:35.880
that allows you to do this um kind of in

00:48:33.200 --> 00:48:38.079
an efficient manner and what this is

00:48:35.880 --> 00:48:41.200
showing is

00:48:38.079 --> 00:48:45.079
um this is the original formatting of a

00:48:41.200 --> 00:48:48.400
prompt that was given I I Believe by uh

00:48:45.079 --> 00:48:50.119
some sort of like uh machine reading or

00:48:48.400 --> 00:48:52.799
document based question answering data

00:48:50.119 --> 00:48:55.480
set which was like passage

00:48:52.799 --> 00:48:58.440
answer if you modify the spacing between

00:48:55.480 --> 00:49:01.680
the the fields that increases your score

00:48:58.440 --> 00:49:04.280
by several percentage points um if you

00:49:01.680 --> 00:49:06.880
remove the colons that increases your

00:49:04.280 --> 00:49:08.720
score by a few more percentage points

00:49:06.880 --> 00:49:10.119
it's kind of silly but like little

00:49:08.720 --> 00:49:11.040
things like this actually can make a

00:49:10.119 --> 00:49:14.240
really big

00:49:11.040 --> 00:49:17.599
difference um if you modify the casing

00:49:14.240 --> 00:49:19.960
this decreases by a lot if you modify

00:49:17.599 --> 00:49:22.440
the casing and remove colons so the

00:49:19.960 --> 00:49:25.200
thing that was useful like adding colons

00:49:22.440 --> 00:49:26.720
here remove colons uh that further

00:49:25.200 --> 00:49:29.280
decrease

00:49:26.720 --> 00:49:31.400
if you forget to add a space between the

00:49:29.280 --> 00:49:32.559
passage and the text that really hurts

00:49:31.400 --> 00:49:35.599
your

00:49:32.559 --> 00:49:38.000
accuracy so this is pretty painful right

00:49:35.599 --> 00:49:40.599
like you don't want to be getting uh

00:49:38.000 --> 00:49:44.160
0.036% accuracy when adding a space

00:49:40.599 --> 00:49:48.680
would give you like 75% accuracy

00:49:44.160 --> 00:49:50.799
right um and one interesting thing is um

00:49:48.680 --> 00:49:53.160
this is looking

00:49:50.799 --> 00:49:56.559
at different

00:49:53.160 --> 00:49:58.520
models and um

00:49:56.559 --> 00:50:00.640
with different models it's pretty

00:49:58.520 --> 00:50:03.599
consistent that many different plausible

00:50:00.640 --> 00:50:05.400
formats that you try the average gives

00:50:03.599 --> 00:50:07.240
you a really low accuracy but there's a

00:50:05.400 --> 00:50:08.760
few outliers that give you really good

00:50:07.240 --> 00:50:11.119
accuracy and these probably correspond

00:50:08.760 --> 00:50:13.400
to the things that it was trained on um

00:50:11.119 --> 00:50:15.880
instruction tuned on or or other things

00:50:13.400 --> 00:50:17.480
like this so number one make sure you're

00:50:15.880 --> 00:50:19.799
using like the canonical prompt

00:50:17.480 --> 00:50:21.240
formatting for the model for sure number

00:50:19.799 --> 00:50:22.640
two you might want to do a little bit of

00:50:21.240 --> 00:50:24.720
additional search to see if you can do

00:50:22.640 --> 00:50:26.960
even better than that so um this is

00:50:24.720 --> 00:50:29.480
something to be very aware

00:50:26.960 --> 00:50:32.480
of

00:50:29.480 --> 00:50:32.480
um

00:50:34.200 --> 00:50:37.680
okay do you have a

00:50:39.599 --> 00:50:43.720
question this is dependent on what it

00:50:41.720 --> 00:50:47.680
sees in trading time another thing

00:50:43.720 --> 00:50:51.920
actually is um this will definitely be

00:50:47.680 --> 00:50:53.200
tighter for uh like a chat GPT or GPT 4

00:50:51.920 --> 00:50:56.599
um because it's been trained on many

00:50:53.200 --> 00:50:59.319
different formats at training time um

00:50:56.599 --> 00:51:00.880
and so the better the model has been

00:50:59.319 --> 00:51:03.520
trained on a lot of different formats

00:51:00.880 --> 00:51:05.559
the less this is going to have an

00:51:03.520 --> 00:51:06.920
effect but you know you're probably not

00:51:05.559 --> 00:51:09.440
going to be retraining a model that

00:51:06.920 --> 00:51:10.799
somebody gives you uh so like this is

00:51:09.440 --> 00:51:12.880
something to be very aware of if you're

00:51:10.799 --> 00:51:14.839
just a downstream newer model especially

00:51:12.880 --> 00:51:17.599
an open source

00:51:14.839 --> 00:51:19.359
model um another thing is how do you

00:51:17.599 --> 00:51:22.280
give instructions to

00:51:19.359 --> 00:51:25.000
models um instructions should be clear

00:51:22.280 --> 00:51:29.280
concise and easy to understand one very

00:51:25.000 --> 00:51:31.559
funny thing is um I think now like

00:51:29.280 --> 00:51:33.280
actually prompting language models is

00:51:31.559 --> 00:51:34.960
very similar to prompting humans

00:51:33.280 --> 00:51:37.119
especially if we're talking about like

00:51:34.960 --> 00:51:38.760
gp4 so if you're not very good at

00:51:37.119 --> 00:51:41.599
explaining things to humans that might

00:51:38.760 --> 00:51:45.440
actually be bad um and you might want to

00:51:41.599 --> 00:51:47.359
practice that and explaining things to

00:51:45.440 --> 00:51:50.319
models might be a good way to practice

00:51:47.359 --> 00:51:51.799
that right so you know um it actually

00:51:50.319 --> 00:51:54.040
can give you feedback without annoying

00:51:51.799 --> 00:51:55.359
your friends by having you explain uh

00:51:54.040 --> 00:51:58.160
things to them in several different ways

00:51:55.359 --> 00:52:00.040
way and seeing how they react so um but

00:51:58.160 --> 00:52:03.680
anyway clear concise easy to understand

00:52:00.040 --> 00:52:05.319
is good um there's this prompting guide

00:52:03.680 --> 00:52:08.599
uh which I I can

00:52:05.319 --> 00:52:13.240
open um this has a prompt engineering

00:52:08.599 --> 00:52:14.520
guide I I I like this site but it it

00:52:13.240 --> 00:52:17.400
does have a bit

00:52:14.520 --> 00:52:18.880
of like variance in the importance of

00:52:17.400 --> 00:52:21.760
the information that tells you but like

00:52:18.880 --> 00:52:23.960
this particular page is nice I feel so

00:52:21.760 --> 00:52:26.160
start simple start with simple

00:52:23.960 --> 00:52:29.520
instructions um

00:52:26.160 --> 00:52:32.119
you should tell the model what it should

00:52:29.520 --> 00:52:36.839
be doing so make sure you say write

00:52:32.119 --> 00:52:39.799
classify summarize translate order um

00:52:36.839 --> 00:52:41.960
and things like this uh it also gives

00:52:39.799 --> 00:52:45.440
some good examples of the level of

00:52:41.960 --> 00:52:47.559
specificity that you should be giving so

00:52:45.440 --> 00:52:49.680
something that's less precise is explain

00:52:47.559 --> 00:52:51.559
the concept of prompt engineering keep

00:52:49.680 --> 00:52:53.920
the explanation short only a few

00:52:51.559 --> 00:52:57.119
sentences and don't be too

00:52:53.920 --> 00:52:58.799
descriptive um it use two to three

00:52:57.119 --> 00:53:00.240
sentences to explain the concept of

00:52:58.799 --> 00:53:02.599
prompt engineering to a high school

00:53:00.240 --> 00:53:04.839
student so what this does is this tells

00:53:02.599 --> 00:53:07.839
you the level of read like the reading

00:53:04.839 --> 00:53:07.839
level

00:53:07.960 --> 00:53:12.520
um so this doesn't even tell you the

00:53:10.200 --> 00:53:14.319
reading level I guess um and then two to

00:53:12.520 --> 00:53:16.240
three sentences is more precise than

00:53:14.319 --> 00:53:19.200
keep it a few sentences don't be too

00:53:16.240 --> 00:53:22.440
descriptive so um the more precise you

00:53:19.200 --> 00:53:25.760
can be the the better it is um one

00:53:22.440 --> 00:53:27.040
interesting thing is like if you ask

00:53:25.760 --> 00:53:28.359
your friend to do something and they

00:53:27.040 --> 00:53:32.400
don't know how to do it they'll complain

00:53:28.359 --> 00:53:34.240
to you but right now uh LMS don't

00:53:32.400 --> 00:53:35.720
complain to you they may in the future

00:53:34.240 --> 00:53:38.680
uh that might be like actually an

00:53:35.720 --> 00:53:40.799
interesting thing to find uh the you

00:53:38.680 --> 00:53:42.319
know interesting methodological thing to

00:53:40.799 --> 00:53:45.240
look at for a project or something like

00:53:42.319 --> 00:53:47.960
that but um right now you need to be

00:53:45.240 --> 00:53:49.040
precise and like there's it doesn't give

00:53:47.960 --> 00:53:51.799
you feedback when you're not being

00:53:49.040 --> 00:53:51.799
precise so you need

00:53:52.000 --> 00:53:56.359
to um separately from this there are

00:53:54.200 --> 00:53:59.160
methods for automatic prompt engineering

00:53:56.359 --> 00:54:00.960
so uh prompt paraphrasing gradient based

00:53:59.160 --> 00:54:02.240
discreet prompt search prompt tuning

00:54:00.960 --> 00:54:06.160
prefix

00:54:02.240 --> 00:54:09.880
tuning so prompt paraphrasing um this is

00:54:06.160 --> 00:54:12.559
a method that uh we proposed a while ago

00:54:09.880 --> 00:54:15.760
um to basically paraphrase an existing

00:54:12.559 --> 00:54:17.280
prompt to get other candidates um it's

00:54:15.760 --> 00:54:19.240
rather simple basically you take a

00:54:17.280 --> 00:54:21.960
prompt you put it through a paraphrasing

00:54:19.240 --> 00:54:24.280
model and it will give you new prompts

00:54:21.960 --> 00:54:25.440
and this is good because it will tend to

00:54:24.280 --> 00:54:28.319
give you things that are natural

00:54:25.440 --> 00:54:29.839
language um you can paraphrase 50 times

00:54:28.319 --> 00:54:32.480
try all of them see which one gives you

00:54:29.839 --> 00:54:37.079
the highest accuracy and then use that

00:54:32.480 --> 00:54:39.280
one um there's also an interesting paper

00:54:37.079 --> 00:54:43.079
uh that demonstrates that you can do

00:54:39.280 --> 00:54:45.240
this iteratively so you paraphrase once

00:54:43.079 --> 00:54:46.599
um you filter down all the candidates

00:54:45.240 --> 00:54:48.119
that do well and then you go in and

00:54:46.599 --> 00:54:49.960
paraphrase them again and you just do

00:54:48.119 --> 00:54:51.960
this over and over again and that can

00:54:49.960 --> 00:54:54.079
give you better results than kind of one

00:54:51.960 --> 00:54:57.079
one off

00:54:54.079 --> 00:54:57.079
paraphrasing

00:54:59.240 --> 00:55:02.079
so that's very simple you can even use a

00:55:01.079 --> 00:55:04.160
large language model to do the

00:55:02.079 --> 00:55:06.599
paraphrasing for you um another thing

00:55:04.160 --> 00:55:08.920
that you can do is gradient based search

00:55:06.599 --> 00:55:11.119
so the way this works is you need to

00:55:08.920 --> 00:55:16.319
have a a model that you can calculate

00:55:11.119 --> 00:55:19.920
gradients for and what you do is you

00:55:16.319 --> 00:55:22.240
calculate you create a seed prompt and

00:55:19.920 --> 00:55:26.000
then you calculate gradients into that

00:55:22.240 --> 00:55:29.760
seed prompt so you treat the

00:55:26.000 --> 00:55:33.160
um you treat each of the tokens here

00:55:29.760 --> 00:55:36.680
like T1 T2 T3 T4

00:55:33.160 --> 00:55:38.240
T5 as their own embeddings you do back

00:55:36.680 --> 00:55:39.920
propop into those embeddings and you

00:55:38.240 --> 00:55:42.799
optimize them to get high accuracy on

00:55:39.920 --> 00:55:44.720
your data set then after you're done

00:55:42.799 --> 00:55:47.319
optimizing them to get high accuracy on

00:55:44.720 --> 00:55:49.079
your data set you clamp them onto the

00:55:47.319 --> 00:55:52.160
nearest neighbor embedding that you

00:55:49.079 --> 00:55:53.520
already have so you basically say okay

00:55:52.160 --> 00:55:56.720
the nearest neighbor to the embedding

00:55:53.520 --> 00:55:58.920
that I learned you um is atmosphere then

00:55:56.720 --> 00:56:02.240
a lot dialogue clone

00:55:58.920 --> 00:56:03.799
totally and so this is this will

00:56:02.240 --> 00:56:05.599
actually give you better results than

00:56:03.799 --> 00:56:07.839
paraphrasing in many cases because the

00:56:05.599 --> 00:56:11.520
search space is less constrained you can

00:56:07.839 --> 00:56:12.960
get these very unnatural prompts uh that

00:56:11.520 --> 00:56:16.280
don't seem to make sense but actually

00:56:12.960 --> 00:56:20.280
work well this has particularly been

00:56:16.280 --> 00:56:22.960
widely used in um adversarial attacks on

00:56:20.280 --> 00:56:25.599
language modals so how can you come up

00:56:22.960 --> 00:56:27.720
with um

00:56:25.599 --> 00:56:31.559
with prompts that

00:56:27.720 --> 00:56:33.319
cause language models to uh do things

00:56:31.559 --> 00:56:36.039
that you don't want them to be

00:56:33.319 --> 00:56:38.920
doing and um there's actually this nice

00:56:36.039 --> 00:56:41.440
paper uh also by people at CMU called

00:56:38.920 --> 00:56:42.960
Universal and transferable adversarial

00:56:41.440 --> 00:56:45.400
attacks on line language

00:56:42.960 --> 00:56:50.559
models and basically what they do is

00:56:45.400 --> 00:56:53.880
they try to optimize the uh they try to

00:56:50.559 --> 00:56:56.839
optimize the prompt to create a prompt

00:56:53.880 --> 00:56:58.599
that causes the model to do bad things

00:56:56.839 --> 00:57:00.039
basically and they try to do it even on

00:56:58.599 --> 00:57:03.440
models that have been trying to not do

00:57:00.039 --> 00:57:05.039
bad things and they demonstrate that

00:57:03.440 --> 00:57:07.359
number one you can cause things like

00:57:05.039 --> 00:57:09.599
models like llama to do bad you know bad

00:57:07.359 --> 00:57:12.559
things like output toxic things tell you

00:57:09.599 --> 00:57:15.599
how to build bombs stuff like that but

00:57:12.559 --> 00:57:18.480
also the same prompts also work on like

00:57:15.599 --> 00:57:22.319
GPD models uh which is kind of like

00:57:18.480 --> 00:57:23.839
interesting and and very uh you know

00:57:22.319 --> 00:57:26.520
confusing in a way because you thought

00:57:23.839 --> 00:57:28.160
this might be explo idiosyncrasies of a

00:57:26.520 --> 00:57:32.440
particular language model but actually

00:57:28.160 --> 00:57:32.440
it's not so I I find this kind of

00:57:33.880 --> 00:57:39.520
fascinating

00:57:36.039 --> 00:57:42.240
so if you take that a step further one

00:57:39.520 --> 00:57:44.079
thing that you can do is you can say oh

00:57:42.240 --> 00:57:46.280
actually there's no reason why we need

00:57:44.079 --> 00:57:48.520
to clamp these embeddings back to an

00:57:46.280 --> 00:57:52.240
existing embedding right so we could

00:57:48.520 --> 00:57:56.079
just optimize the prompts the embeddings

00:57:52.240 --> 00:57:57.720
of the prompts that go for a task and

00:57:56.079 --> 00:58:02.000
not clamp them back to embeddings and

00:57:57.720 --> 00:58:03.599
just keep them as is so um what I mean

00:58:02.000 --> 00:58:07.079
by that is like right here it's

00:58:03.599 --> 00:58:09.160
optimizing T1 T2 T3 T4 T5 and then

00:58:07.079 --> 00:58:11.359
clamping that back to Atmosphere a lot

00:58:09.160 --> 00:58:13.960
dialog clone totally but just keep them

00:58:11.359 --> 00:58:16.160
as is and don't worry about them like

00:58:13.960 --> 00:58:18.039
actually being a token in the model

00:58:16.160 --> 00:58:19.400
because if you have control over your

00:58:18.039 --> 00:58:21.200
model you can just add them as new

00:58:19.400 --> 00:58:25.960
elements in the vocabulary and you're

00:58:21.200 --> 00:58:28.440
fine right so what they demonstrate in

00:58:25.960 --> 00:58:31.520
this paper is that instead of taking

00:58:28.440 --> 00:58:33.440
your 11 billion parameter model and

00:58:31.520 --> 00:58:35.920
training the whole 11 billion parameter

00:58:33.440 --> 00:58:38.359
model for many different tasks on many

00:58:35.920 --> 00:58:40.079
different data sets they just train

00:58:38.359 --> 00:58:42.039
these prompts which are like 20K

00:58:40.079 --> 00:58:44.039
parameters each I I forget how long it

00:58:42.039 --> 00:58:46.280
is it's like 10 tokens or 20 tokens or

00:58:44.039 --> 00:58:48.079
something like that um and train it on

00:58:46.280 --> 00:58:49.640
all of the the data sets here and you

00:58:48.079 --> 00:58:50.680
don't actually need to do multitask

00:58:49.640 --> 00:58:52.200
learning you don't need to train on

00:58:50.680 --> 00:58:53.720
multiple tasks at the same time you can

00:58:52.200 --> 00:58:56.119
just train on a single

00:58:53.720 --> 00:58:58.599
task

00:58:56.119 --> 00:59:01.000
so now let's take that even a step

00:58:58.599 --> 00:59:03.640
further so this is only training the

00:59:01.000 --> 00:59:06.359
embeddings that you input into the model

00:59:03.640 --> 00:59:08.160
there's a method called prefix tuning

00:59:06.359 --> 00:59:10.319
and the way prefix tuning works is

00:59:08.160 --> 00:59:12.280
instead of training only the embeddings

00:59:10.319 --> 00:59:14.799
that go into the model they actually

00:59:12.280 --> 00:59:18.920
train a prefix that you then append to

00:59:14.799 --> 00:59:20.839
every layer of the model so prompt

00:59:18.920 --> 00:59:23.319
tuning basically does this for the first

00:59:20.839 --> 00:59:24.839
layer of the model prefix tuning does

00:59:23.319 --> 00:59:28.400
this for every layer of the model you

00:59:24.839 --> 00:59:30.319
append a prefix uh for every day so it's

00:59:28.400 --> 00:59:32.200
just a more expressive version of

00:59:30.319 --> 00:59:36.119
prompting

00:59:32.200 --> 00:59:40.200
essentially so these are all kinds of

00:59:36.119 --> 00:59:43.680
gradual steps from a human created

00:59:40.200 --> 00:59:47.880
prompt into something that is basically

00:59:43.680 --> 00:59:50.839
training a a prompt or a prefix to the

00:59:47.880 --> 00:59:52.960
model so I I would take questions but

00:59:50.839 --> 00:59:55.200
let me get to the end of this section uh

00:59:52.960 --> 00:59:58.839
also because uh I think there's

00:59:55.200 --> 01:00:00.720
interesting analogies here so in the

00:59:58.839 --> 01:00:02.880
next class I'm going to talk about

01:00:00.720 --> 01:00:04.440
parameter efficient fine-tuning methods

01:00:02.880 --> 01:00:06.960
which is kind of a more

01:00:04.440 --> 01:00:10.000
General it's a more

01:00:06.960 --> 01:00:11.480
General version of prompt tuning or

01:00:10.000 --> 01:00:13.280
prefix tuning there are methods that

01:00:11.480 --> 01:00:15.960
tune a small number of parameters to get

01:00:13.280 --> 01:00:17.400
the model to do something and there's a

01:00:15.960 --> 01:00:18.880
bunch of different parameter efficient

01:00:17.400 --> 01:00:21.520
tuning methods many people may have

01:00:18.880 --> 01:00:23.880
heard of something like Laura uh or

01:00:21.520 --> 01:00:25.440
adapters um I just talked about prefix

01:00:23.880 --> 01:00:28.119
tuning

01:00:25.440 --> 01:00:30.960
so essentially prompt tuning and prefix

01:00:28.119 --> 01:00:33.359
tuning are part of this more General

01:00:30.960 --> 01:00:36.680
class of parameter efficient find tuning

01:00:33.359 --> 01:00:39.240
methods and so what we can say is

01:00:36.680 --> 01:00:41.119
actually prompting is fine-tuning

01:00:39.240 --> 01:00:42.920
prompting is a way of fine-tuning the

01:00:41.119 --> 01:00:46.799
model or getting the model to perform a

01:00:42.920 --> 01:00:49.839
particular task um and we have this

01:00:46.799 --> 01:00:53.720
taxonomy of we have prompts in natural

01:00:49.839 --> 01:00:55.160
language that are created uh by humans

01:00:53.720 --> 01:00:57.240
actually maybe I should say manual

01:00:55.160 --> 01:00:59.559
prompt engineering here this was first

01:00:57.240 --> 01:01:01.480
done in the gpd2 paper where they

01:00:59.559 --> 01:01:04.359
demonstrate that models uh models could

01:01:01.480 --> 01:01:06.200
solve tasks by doing it this way prompt

01:01:04.359 --> 01:01:07.760
paraphrasing is a step up from this

01:01:06.200 --> 01:01:09.799
because it's no longer relying on human

01:01:07.760 --> 01:01:12.680
engineering and you can you know expand

01:01:09.799 --> 01:01:15.280
to a broader set of prompts um it can

01:01:12.680 --> 01:01:17.359
always start with human created prompts

01:01:15.280 --> 01:01:20.240
so it's kind of like broader uh than

01:01:17.359 --> 01:01:21.799
that discrete prompt search doesn't

01:01:20.240 --> 01:01:23.599
necessarily need to rely on a

01:01:21.799 --> 01:01:25.559
paraphrasing model it could rely on like

01:01:23.599 --> 01:01:26.760
gradient-based models or something else

01:01:25.559 --> 01:01:29.240
like that to give you something that's

01:01:26.760 --> 01:01:32.559
not actually natural language uh kind of

01:01:29.240 --> 01:01:35.920
just random tokens continuous prompts or

01:01:32.559 --> 01:01:38.119
prompt tuning is a step above that

01:01:35.920 --> 01:01:41.039
multi-layer continuous prompts or prefix

01:01:38.119 --> 01:01:42.520
tuning is a layer above that parameter

01:01:41.039 --> 01:01:43.520
efficient tuning is more General than

01:01:42.520 --> 01:01:45.359
that and then you have all training

01:01:43.520 --> 01:01:49.160
methods so including fine tuning your

01:01:45.359 --> 01:01:52.680
model and so what are the implications

01:01:49.160 --> 01:01:55.760
of this um I think so a lot of people

01:01:52.680 --> 01:01:58.720
when prompting came out they were like

01:01:55.760 --> 01:02:00.640
prompting methods are very hacky I don't

01:01:58.720 --> 01:02:03.839
like how we have to do manual prompt

01:02:00.640 --> 01:02:08.160
engineering um it seems like a dark art

01:02:03.839 --> 01:02:11.000
as opposed to like you know actually you

01:02:08.160 --> 01:02:14.160
know some sort of well understood

01:02:11.000 --> 01:02:16.839
fine-tuning method that we could use um

01:02:14.160 --> 01:02:20.520
but I I actually like them I like

01:02:16.839 --> 01:02:23.920
prompting a lot because um if anybody is

01:02:20.520 --> 01:02:25.960
familiar with like basian basian

01:02:23.920 --> 01:02:27.920
statistics or machine learning we have

01:02:25.960 --> 01:02:28.799
the concept of like a prior probability

01:02:27.920 --> 01:02:31.200
over

01:02:28.799 --> 01:02:32.359
parameters and then a probability that

01:02:31.200 --> 01:02:34.680
we get

01:02:32.359 --> 01:02:37.880
after after fine tuning the model or

01:02:34.680 --> 01:02:40.440
after training the model and prompts in

01:02:37.880 --> 01:02:42.640
a way are our first like good prior over

01:02:40.440 --> 01:02:43.880
neural network models they give us the

01:02:42.640 --> 01:02:46.319
ability to

01:02:43.880 --> 01:02:48.559
specify what task the model should be

01:02:46.319 --> 01:02:51.880
doing or like a general idea of what

01:02:48.559 --> 01:02:54.200
task the model should be doing before we

01:02:51.880 --> 01:02:56.359
ask the model to actually do the task

01:02:54.200 --> 01:02:58.640
and and so we can either use that prior

01:02:56.359 --> 01:03:02.119
Asis we can use a prompted model Asis

01:02:58.640 --> 01:03:04.839
without doing any additional tuning or

01:03:02.119 --> 01:03:06.480
we could take the prior that we have

01:03:04.839 --> 01:03:07.920
given to the model by using a natural

01:03:06.480 --> 01:03:09.039
language description of the task it

01:03:07.920 --> 01:03:12.079
should be

01:03:09.039 --> 01:03:14.799
doing and then combine it with fineing

01:03:12.079 --> 01:03:17.039
so we can take the prompted

01:03:14.799 --> 01:03:19.279
model we can

01:03:17.039 --> 01:03:21.640
initialize we can initialize the

01:03:19.279 --> 01:03:23.960
distribution of this like Cas a prompt

01:03:21.640 --> 01:03:25.720
using the prompt using a human created

01:03:23.960 --> 01:03:28.160
prompt and then go on and fine-tune it

01:03:25.720 --> 01:03:30.960
on lots of training data as well and

01:03:28.160 --> 01:03:33.799
there's a method for doing that um by

01:03:30.960 --> 01:03:35.880
shik and schutza uh called uh pattern

01:03:33.799 --> 01:03:37.559
exploiting training where they do

01:03:35.880 --> 01:03:39.799
exactly that they basically initialize

01:03:37.559 --> 01:03:41.720
with a manually created prompt and then

01:03:39.799 --> 01:03:44.559
they find the model on finding inator

01:03:41.720 --> 01:03:46.400
after that so um that's a reason why I

01:03:44.559 --> 01:03:47.920
like prompting based methods they they

01:03:46.400 --> 01:03:49.720
give us this like really nice way to

01:03:47.920 --> 01:03:53.039
very quickly create a system but we can

01:03:49.720 --> 01:03:56.079
also have you know whatever level of

01:03:53.039 --> 01:03:59.880
additional training on top of that

01:03:56.079 --> 01:03:59.880
cool so that's a little bit early I'm