prompt_eng_06_transforming_v3.srt

1
00:00:05,133 --> 00:00:08,033
Las Vegas models are very good
at transforming

2
00:00:08,033 --> 00:00:12,300
as input to a different format, such as
inputting a piece of text in one language

3
00:00:12,600 --> 00:00:16,500
and transforming it,
or translating it to a different language

4
00:00:16,766 --> 00:00:19,200
or helping with spelling
and grammar correction.

5
00:00:19,200 --> 00:00:21,000
So taking this input, the piece of text

6
00:00:21,000 --> 00:00:24,666
that may not be fully grammatical
and helping you to fix that up a bit,

7
00:00:25,166 --> 00:00:30,900
or even transforming formats
such as inputting HTML and opposing JSON.

8
00:00:31,200 --> 00:00:34,933
So there's a bunch of applications
that I used to write somewhat painfully

9
00:00:34,933 --> 00:00:37,300
with a bunch of regular expressions
that would definitely

10
00:00:37,333 --> 00:00:40,100
be much more simply implemented
now with a large language.

11
00:00:40,100 --> 00:00:42,033
And although in a few prompts,

12
00:00:42,033 --> 00:00:46,366
yeah, I used to GPT to proofread
pretty much everything I write these days.

13
00:00:46,533 --> 00:00:49,566
So I'm excited to show you
some more examples in the notebook now.

14
00:00:51,000 --> 00:00:53,033
So first we'll

15
00:00:53,033 --> 00:00:57,566
import
openai and also use the same completion

16
00:00:57,566 --> 00:01:00,600
helper function that we've been using
throughout the videos.

17
00:01:01,233 --> 00:01:04,233
And the first thing we'll do
is a translation task.

18
00:01:04,600 --> 00:01:07,900
So large language
models are trained on a lot of text

19
00:01:07,900 --> 00:01:10,500
from kind of many sources,
a lot of which is the Internet.

20
00:01:10,500 --> 00:01:13,000
And this is kind of,
of course, in many different languages.

21
00:01:13,000 --> 00:01:16,466
So this kind of imbues the model
with the ability to do translation.

22
00:01:16,866 --> 00:01:19,200
And these models know
kind of hundreds of languages

23
00:01:19,433 --> 00:01:21,633
to varying degrees of proficiency.

24
00:01:21,633 --> 00:01:24,900
And so we'll go through some examples
of how to use this capability.

25
00:01:26,400 --> 00:01:29,100
So let's start off with something simple.

26
00:01:29,100 --> 00:01:30,600
So in this first example,

27
00:01:30,600 --> 00:01:34,366
the prompt is translate
the following English text to Spanish.

28
00:01:35,100 --> 00:01:37,400
Hi, I would like to order a blender

29
00:01:37,400 --> 00:01:40,900
and the response is allow me
goose diarrhea

30
00:01:41,166 --> 00:01:44,400
or then a liquid quadra.

31
00:01:44,400 --> 00:01:47,733
And I'm very sorry to all of you
Spanish speakers.

32
00:01:47,733 --> 00:01:51,700
I never learn Spanish,
unfortunately, as you can definitely tell.

33
00:01:53,400 --> 00:01:55,900
Okay, let's try another example.

34
00:01:56,566 --> 00:01:59,500
So in this example,

35
00:01:59,533 --> 00:02:02,566
the prompt is
tell me what language this is.

36
00:02:03,800 --> 00:02:05,833
And then this is in French

37
00:02:06,166 --> 00:02:08,400
Colombian food lump of death

38
00:02:09,000 --> 00:02:12,133
and say let's from this.

39
00:02:12,133 --> 00:02:17,800
And the model has identified
that this is French.

40
00:02:17,800 --> 00:02:20,833
The model can also do
multiple translations at once.

41
00:02:21,666 --> 00:02:24,000
So in this example, let's say translate

42
00:02:24,000 --> 00:02:27,266
the following text to French and Spanish.

43
00:02:27,266 --> 00:02:27,933
And you know what?

44
00:02:27,933 --> 00:02:30,000
Let's add another and English

45
00:02:31,366 --> 00:02:34,466
pirates.

46
00:02:34,466 --> 00:02:40,533
The text is I want to order a basketball.

47
00:02:40,533 --> 00:02:43,233
So here we have French, Spanish

48
00:02:44,100 --> 00:02:46,433
and English pirates.

49
00:02:46,966 --> 00:02:49,666
So in some languages
the translation can change

50
00:02:49,666 --> 00:02:52,433
depending on the speaker's relationship
to the listener.

51
00:02:52,800 --> 00:02:55,933
And you can also explain this
to the language model,

52
00:02:55,933 --> 00:02:59,166
and it will be able to kind of
translate it accordingly.

53
00:02:59,633 --> 00:03:03,033
So in this example,
we say translate the following text

54
00:03:03,033 --> 00:03:05,833
to Spanish
in both the formal and informal forms.

55
00:03:06,166 --> 00:03:09,166
Would you like to order a pillow
and also notice here

56
00:03:09,333 --> 00:03:12,866
we're using a different delimiter
than these back texts.

57
00:03:12,900 --> 00:03:15,900
It doesn't really matter as long as it's
kind of a clear separation.

58
00:03:17,733 --> 00:03:20,733
So here we have the formal and informal.

59
00:03:21,733 --> 00:03:24,933
So informal as when you're speaking
to someone who's maybe senior to you

60
00:03:24,933 --> 00:03:28,333
or you're in a professional situation,
that's when you use a formal tone.

61
00:03:28,333 --> 00:03:32,166
And then informal is when you're speaking
to maybe a group of friends.

62
00:03:32,766 --> 00:03:38,033
I don't actually speak Spanish, but my dad
does, and he says that this is correct.

63
00:03:38,033 --> 00:03:41,466
So for the next example,
we're going to pretend

64
00:03:41,466 --> 00:03:45,300
that we're in charge
of a multinational e-commerce company.

65
00:03:45,466 --> 00:03:48,833
And so the user messages are going to be
in all different languages.

66
00:03:48,833 --> 00:03:50,500
And so users are going to be telling us

67
00:03:50,500 --> 00:03:54,000
about the issues
in a wide variety of languages.

68
00:03:54,700 --> 00:03:57,666
So we need a universal translator.

69
00:03:57,666 --> 00:03:59,200
So first we'll just paste

70
00:03:59,200 --> 00:04:03,066
in a list of user messages
in a variety of different languages.

71
00:04:04,700 --> 00:04:06,900
And now we will loop through

72
00:04:10,466 --> 00:04:12,700
each of these user messages.

73
00:04:12,700 --> 00:04:17,966
So for issue and user messages,

74
00:04:17,966 --> 00:04:21,466
and then I'm going to copy
over the slightly longer code block.

75
00:04:22,900 --> 00:04:25,800
And so the first thing we'll do
is ask the model

76
00:04:25,800 --> 00:04:29,366
to tell us what language the issue is in.

77
00:04:29,500 --> 00:04:31,500
So here's the prompt.

78
00:04:31,500 --> 00:04:36,600
Then we'll print out the original
messages, language and the issue,

79
00:04:37,100 --> 00:04:42,600
and then we'll ask the model
to translate it into English and Korean.

80
00:04:42,600 --> 00:04:45,500
So that's from this.

81
00:04:45,500 --> 00:04:54,800
So the original message in French,
so we have a variety of languages

82
00:04:54,800 --> 00:04:58,800
and then the model translates them
into English and then Korean

83
00:04:59,466 --> 00:05:01,033
and you can kind of see here.

84
00:05:01,033 --> 00:05:03,900
So the model says this is French.

85
00:05:04,400 --> 00:05:09,633
So that's because the response from this
prompt is going to be, this is French.

86
00:05:10,000 --> 00:05:12,566
You could try editing this prompt
to say something like,

87
00:05:12,766 --> 00:05:14,066
Tell me what language this is.

88
00:05:14,066 --> 00:05:16,166
Respond with only one

89
00:05:16,166 --> 00:05:19,466
with only one word, or
don't use a sentence, that kind of thing.

90
00:05:19,466 --> 00:05:23,733
If you wanted this to just be one word,
or you could ask for it in a Jason format

91
00:05:23,733 --> 00:05:24,466
or something like that,

92
00:05:24,466 --> 00:05:29,066
which would probably encourage it
to not use a whole sentence.

93
00:05:29,066 --> 00:05:30,800
And so amazing.

94
00:05:30,800 --> 00:05:33,500
You've just built a universal translator.

95
00:05:33,800 --> 00:05:36,066
And also feel free to pause the video

96
00:05:36,433 --> 00:05:40,266
and add kind of any other languages
you want to try here.

97
00:05:40,266 --> 00:05:45,900
Maybe languages you speak your self
and see how the model does.

98
00:05:45,900 --> 00:05:49,833
So the next thing we're going to dive
into is tone transformation.

99
00:05:50,300 --> 00:05:53,033
Writing can vary
based on an intended audience.

100
00:05:53,033 --> 00:05:58,000
You know the way that I would write
an email to a colleague or a professor

101
00:05:58,000 --> 00:06:01,200
is obviously going to be quite different
to the way I text my younger brother.

102
00:06:01,800 --> 00:06:06,600
And so interactivity can actually
also help produce different tones.

103
00:06:06,600 --> 00:06:08,833
So let's look at some examples.

104
00:06:09,333 --> 00:06:11,800
So in this first example,

105
00:06:11,800 --> 00:06:15,400
the prompt is translate the following
from slang to a business letter.

106
00:06:15,800 --> 00:06:16,966
Dude, this is Joe.

107
00:06:16,966 --> 00:06:20,966
Check out the spec on the standing lamp.

108
00:06:20,966 --> 00:06:26,233
So let's execute this.

109
00:06:26,233 --> 00:06:30,000
And as you can see,
we have a much more formal business letter

110
00:06:30,600 --> 00:06:34,933
with a proposal for a standing lamp
specification.

111
00:06:34,933 --> 00:06:39,266
The next thing that we're going to do
is to convert between different formats.

112
00:06:39,800 --> 00:06:43,566
GBC is very good at translating
between different formats such as JSON to

113
00:06:43,800 --> 00:06:47,366
HTML, XML, all kinds of things, markdown.

114
00:06:48,666 --> 00:06:49,466
And so in the

115
00:06:49,466 --> 00:06:52,200
prompt will describe both the input
and the output formats.

116
00:06:52,666 --> 00:06:55,466
So here is an example.

117
00:06:55,833 --> 00:06:59,700
So we have this JSON that contains

118
00:07:00,300 --> 00:07:04,433
a list of restaurant employees
with their name and email.

119
00:07:04,433 --> 00:07:08,400
And then in the prompt,
we're going to ask the model

120
00:07:08,400 --> 00:07:11,500
to translate this from JSON to its HTML.

121
00:07:11,800 --> 00:07:15,500
So the prompt is translate
the following Python dictionary from JSON

122
00:07:15,733 --> 00:07:21,600
to an HTML table
with column headers and titles,

123
00:07:21,600 --> 00:07:25,800
and then we'll get the response

124
00:07:26,600 --> 00:07:29,933
from the model and print it.

125
00:07:29,933 --> 00:07:33,300
So here we have some HTML displaying

126
00:07:33,300 --> 00:07:36,933
all of the employee names and emails.

127
00:07:38,533 --> 00:07:40,133
And so now let's

128
00:07:40,133 --> 00:07:42,600
see if we can actually view this HTML.

129
00:07:43,366 --> 00:07:48,033
So we're going to use this
display function from this Python library

130
00:07:48,433 --> 00:07:54,533
display HTML response.

131
00:07:54,533 --> 00:07:57,900
And here you can see that
this is a properly formatted

132
00:07:58,433 --> 00:08:01,233
HTML table.

133
00:08:01,233 --> 00:08:06,600
The next transformation task we're going
to do is spell check and grammar checking.

134
00:08:07,033 --> 00:08:10,566
And this is a really kind of popular
use for GPT.

135
00:08:10,566 --> 00:08:12,633
I highly recommend doing this.

136
00:08:12,633 --> 00:08:15,066
I do this all the time
and it's especially useful

137
00:08:15,066 --> 00:08:17,166
when you're walking
in a non-native language.

138
00:08:17,166 --> 00:08:20,400
And so here are some examples
of some common grammar

139
00:08:20,400 --> 00:08:24,600
and spelling problems and how the language
model can help address these.

140
00:08:25,500 --> 00:08:28,866
So I'm going to paste in a list of

141
00:08:29,133 --> 00:08:32,433
sentences that have some grammatical
or spelling errors,

142
00:08:33,400 --> 00:08:35,400
and then we're going to

143
00:08:35,400 --> 00:08:37,933
loop through each of these sentences

144
00:08:42,100 --> 00:08:46,800
and ask the model to proofread these

145
00:08:49,200 --> 00:08:51,366
proofread and correct,

146
00:08:52,033 --> 00:08:54,533
and then we'll use some delimiter limiters

147
00:09:04,333 --> 00:09:06,400
and then we will get the response

148
00:09:06,400 --> 00:09:27,933
and print it as usual.

149
00:09:27,933 --> 00:09:30,866
And so the model is able to correct

150
00:09:31,333 --> 00:09:33,766
all of these grammatical errors,

151
00:09:34,066 --> 00:09:36,600
and we could use some of the techniques
that we've discussed before.

152
00:09:36,833 --> 00:09:39,633
So we could to improve the prompt.

153
00:09:39,633 --> 00:09:42,600
We can say proofreading, correct
the following text

154
00:09:46,100 --> 00:09:47,900
and rewrite

155
00:09:56,433 --> 00:09:59,733
and rewrite the whole

156
00:10:02,366 --> 00:10:04,500
and rewrite it,

157
00:10:04,800 --> 00:10:06,366
corrected

158
00:10:08,733 --> 00:10:09,300
version.

159
00:10:09,300 --> 00:10:13,700
If you don't find any errors,

160
00:10:14,800 --> 00:10:16,533
just say

161
00:10:16,900 --> 00:10:17,766
no errors.

162
00:10:17,766 --> 00:10:21,266
Found.

163
00:10:21,266 --> 00:10:26,466
Let's try this.

164
00:10:26,466 --> 00:10:30,166
So this way
we were able to still using quotes here,

165
00:10:31,233 --> 00:10:33,633
but you can imagine
you'd be able to find a way

166
00:10:33,633 --> 00:10:37,366
with a little bit of iterative
prompt development

167
00:10:37,633 --> 00:10:40,700
to kind of find a problem that works
more reliably every single time.

168
00:10:41,966 --> 00:10:43,833
And so now we'll do another example.

169
00:10:43,833 --> 00:10:48,033
It's always useful to check your text
before you post it in a public forum.

170
00:10:48,233 --> 00:10:50,700
And so we'll go through an example
of checking a review.

171
00:10:51,633 --> 00:10:54,233
And so here is a review
about a stuffed panda.

172
00:10:54,233 --> 00:11:01,100
And so we're going to ask the model
to proofread and correct the review.

173
00:11:01,100 --> 00:11:01,400
Great.

174
00:11:01,400 --> 00:11:06,733
So we have this corrected version, and one
cool thing we can do is find the kind

175
00:11:06,733 --> 00:11:10,500
of differences between our original review
and the model's output.

176
00:11:11,000 --> 00:11:15,033
So we're going to use this red lines
Python package to do this,

177
00:11:15,033 --> 00:11:19,066
and we're going to get the diff
between the original

178
00:11:20,133 --> 00:11:22,900
text of our review and the model output

179
00:11:23,200 --> 00:11:27,366
and then display this.

180
00:11:27,366 --> 00:11:31,500
And so here you can see the difference
between the original review and

181
00:11:31,866 --> 00:11:35,033
the model output and the kind of things
that have been corrected.

182
00:11:35,666 --> 00:11:40,766
So the prompt that we use was proofread
and correct this review, but you can also

183
00:11:40,766 --> 00:11:45,633
make kind of more dramatic changes,
changes to tone and that kind of thing.

184
00:11:45,933 --> 00:11:50,133
So let's try one more thing.

185
00:11:50,533 --> 00:11:54,100
So in this prompt,
we're going to ask the model to proofread

186
00:11:54,100 --> 00:11:57,766
and correct this same review,
but also make it more compelling

187
00:11:58,033 --> 00:12:01,733
and ensure that it follows
APA style and targets an advanced reader.

188
00:12:02,033 --> 00:12:04,766
And we're also going to ask
for the output in markdown format.

189
00:12:05,033 --> 00:12:08,166
And so we're using the same text
from the original review up here.

190
00:12:09,000 --> 00:12:13,433
So let's execute this.

191
00:12:13,433 --> 00:12:16,100
And here we have a expanded

192
00:12:16,633 --> 00:12:19,533
APA style review of the soft panda.

193
00:12:20,566 --> 00:12:22,466
So this is it for the transforming video.

194
00:12:22,466 --> 00:12:25,900
Next up, we have expanding where
we'll take a shorter prompt

195
00:12:25,900 --> 00:12:30,966
and kind of generate a longer, more free
form response from a language model.