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

1
00:00:01,319 --> 00:00:07,560
um today I want to talk about prompting

2
00:00:03,919 --> 00:00:09,639
and uh prompting is kind of a new uh

3
00:00:07,560 --> 00:00:11,320
Paradigm as of a few years ago with

4
00:00:09,639 --> 00:00:15,120
interacting with models it's now kind of

5
00:00:11,320 --> 00:00:16,880
the standard uh in doing so and

6
00:00:15,120 --> 00:00:19,880
basically what we do is we encourage a

7
00:00:16,880 --> 00:00:21,840
pre-trained model to make predictions by

8
00:00:19,880 --> 00:00:24,039
providing a textual prompt specifying

9
00:00:21,840 --> 00:00:25,960
the task to be done this is how you

10
00:00:24,039 --> 00:00:28,960
always interact with chat GPT or

11
00:00:25,960 --> 00:00:33,200
anything else like this

12
00:00:28,960 --> 00:00:36,200
um so prompting fundamentals uh the way

13
00:00:33,200 --> 00:00:38,360
that basic prompting works is you append

14
00:00:36,200 --> 00:00:42,079
a textual string to the beginning of the

15
00:00:38,360 --> 00:00:44,079
output and you complete it and exactly

16
00:00:42,079 --> 00:00:45,800
how you complete it can be based on any

17
00:00:44,079 --> 00:00:48,800
of the generation methods that we talked

18
00:00:45,800 --> 00:00:51,559
about in the previous class uh you know

19
00:00:48,800 --> 00:00:55,160
beam search it can be uh sampling it can

20
00:00:51,559 --> 00:00:58,480
be MBR or self-consistency or whatever

21
00:00:55,160 --> 00:01:00,960
else um so I I put in when a dog sees a

22
00:00:58,480 --> 00:01:03,680
squirrel it will usually

23
00:01:00,960 --> 00:01:06,280
um into gpt2 small which is a very small

24
00:01:03,680 --> 00:01:08,960
language model says Be Afraid of

25
00:01:06,280 --> 00:01:10,560
Anything unusual as an exception that's

26
00:01:08,960 --> 00:01:13,720
when a squirrel is usually afraid to

27
00:01:10,560 --> 00:01:16,280
bitee um so as you can see if the model

28
00:01:13,720 --> 00:01:19,560
is not super great you get a kind of not

29
00:01:16,280 --> 00:01:24,119
very great response also um but then I

30
00:01:19,560 --> 00:01:25,960
CED it into gp2 XL and uh what it says

31
00:01:24,119 --> 00:01:28,159
when a dog sees a squirrel it will

32
00:01:25,960 --> 00:01:30,640
usually lick the squirrel it will also

33
00:01:28,159 --> 00:01:34,000
touch its nose to the squirrel the tail

34
00:01:30,640 --> 00:01:37,880
and nose if it can um which might be

35
00:01:34,000 --> 00:01:40,280
true um one thing I I should note is

36
00:01:37,880 --> 00:01:43,040
when I generated these I used uh like

37
00:01:40,280 --> 00:01:45,200
actual regular ancestral sampling so I

38
00:01:43,040 --> 00:01:47,159
set the temperature to one I didn't do

39
00:01:45,200 --> 00:01:49,600
top feed didn't do top K or anything

40
00:01:47,159 --> 00:01:51,040
like this so this is a raw view of what

41
00:01:49,600 --> 00:01:53,799
the language model thinks is like

42
00:01:51,040 --> 00:01:58,479
actually a reasonable answer um if I

43
00:01:53,799 --> 00:02:00,159
modified the code to do something else

44
00:01:58,479 --> 00:02:02,560
actually maybe I can I can do that that

45
00:02:00,159 --> 00:02:04,960
right now but if I modified the code to

46
00:02:02,560 --> 00:02:08,879
use a

47
00:02:04,960 --> 00:02:12,119
different output we can actually see uh

48
00:02:08,879 --> 00:02:12,119
the different result that we

49
00:02:13,599 --> 00:02:17,959
get since I I have it here

50
00:02:18,360 --> 00:02:23,879
anyway actually sorry I'll need to

51
00:02:20,360 --> 00:02:27,239
modify the code on my my screen here

52
00:02:23,879 --> 00:02:32,120
um so I will

53
00:02:27,239 --> 00:02:35,040
set uh top K to 50 top P to

54
00:02:32,120 --> 00:02:38,360
0.95 so you see I I changed the

55
00:02:35,040 --> 00:02:38,360
generation parameters

56
00:02:38,760 --> 00:02:46,400
here and I'll uh run all of

57
00:02:43,159 --> 00:02:50,319
them you can see the uh the result that

58
00:02:46,400 --> 00:02:51,840
we get in a little bit but basically um

59
00:02:50,319 --> 00:02:54,800
so this is the standard method for

60
00:02:51,840 --> 00:02:57,319
prompting I intentionally use gpt2 small

61
00:02:54,800 --> 00:02:58,800
and gpt2 XL here because these are raw

62
00:02:57,319 --> 00:03:01,879
based language models they were just

63
00:02:58,800 --> 00:03:05,440
pre-trained as language models and so

64
00:03:01,879 --> 00:03:06,920
when we prompt them we're getting a

65
00:03:05,440 --> 00:03:09,200
language model that was just trained on

66
00:03:06,920 --> 00:03:12,280
lots of texts view of what is likely

67
00:03:09,200 --> 00:03:13,760
next text um there are other ways to

68
00:03:12,280 --> 00:03:15,599
train language models like instruction

69
00:03:13,760 --> 00:03:18,040
tuning and rlf which I'm going to be

70
00:03:15,599 --> 00:03:19,480
talking in future classes and if that's

71
00:03:18,040 --> 00:03:21,760
the case you might get a different

72
00:03:19,480 --> 00:03:23,159
response here so when a dog sees a

73
00:03:21,760 --> 00:03:25,720
squirrel it will usually get angry

74
00:03:23,159 --> 00:03:27,319
scratched the squirrel and run off uh

75
00:03:25,720 --> 00:03:29,080
some dogs may also attempt to capture

76
00:03:27,319 --> 00:03:30,799
the squirrel or attempt to eat it dogs

77
00:03:29,080 --> 00:03:32,599
will often to pick up the squirrel and

78
00:03:30,799 --> 00:03:36,400
eat it

79
00:03:32,599 --> 00:03:40,680
for it was more uh more violent than I

80
00:03:36,400 --> 00:03:44,280
expected any um

81
00:03:40,680 --> 00:03:45,720
so but anyway I think that like actually

82
00:03:44,280 --> 00:03:47,080
you can see that when I used the

83
00:03:45,720 --> 00:03:48,920
different generation parameters it

84
00:03:47,080 --> 00:03:51,480
actually gave me something that was

85
00:03:48,920 --> 00:03:54,319
maybe more typical than lick so lick is

86
00:03:51,480 --> 00:03:56,840
maybe a kind of unusual uh answer here

87
00:03:54,319 --> 00:03:58,680
but anyway

88
00:03:56,840 --> 00:04:03,040
cool

89
00:03:58,680 --> 00:04:05,680
so that's the basic idea of prompting we

90
00:04:03,040 --> 00:04:08,480
tend to use prompting to try to solve

91
00:04:05,680 --> 00:04:10,680
problems also so it's not just to

92
00:04:08,480 --> 00:04:14,200
complete text although completing text

93
00:04:10,680 --> 00:04:17,320
is useful and important like I complete

94
00:04:14,200 --> 00:04:19,199
text in my Gmail all the time uh you

95
00:04:17,320 --> 00:04:20,600
know it it's constantly giving me

96
00:04:19,199 --> 00:04:23,440
suggestions about what I should write

97
00:04:20,600 --> 00:04:24,800
next and I do tab autoc complete um you

98
00:04:23,440 --> 00:04:28,040
know on your phone you're doing that

99
00:04:24,800 --> 00:04:29,919
that's also using a language model um

100
00:04:28,040 --> 00:04:32,320
but very often we'll use prompting to do

101
00:04:29,919 --> 00:04:34,440
things other than just completing Texs

102
00:04:32,320 --> 00:04:36,000
and when we do this uh this is kind of

103
00:04:34,440 --> 00:04:38,199
the standard workflow for how we solve

104
00:04:36,000 --> 00:04:41,280
NLP tasks with prompting the way we do

105
00:04:38,199 --> 00:04:43,360
this is we fill in a prompt template

106
00:04:41,280 --> 00:04:46,080
predict the answer and post-process the

107
00:04:43,360 --> 00:04:46,080
answer in some

108
00:04:46,320 --> 00:04:51,880
way so prompt templates are templates

109
00:04:49,280 --> 00:04:55,280
where you will actually uh that you will

110
00:04:51,880 --> 00:04:57,479
fill in with an actual input and so if

111
00:04:55,280 --> 00:05:00,479
we have an input X which is something

112
00:04:57,479 --> 00:05:04,880
like I love this movie our template will

113
00:05:00,479 --> 00:05:08,360
be something like X overall it was Z or

114
00:05:04,880 --> 00:05:10,680
overall it was and so if we do that when

115
00:05:08,360 --> 00:05:13,320
we actually want to make a prediction we

116
00:05:10,680 --> 00:05:14,840
will uh convert this into the actual

117
00:05:13,320 --> 00:05:16,880
prompt we feed into the language model

118
00:05:14,840 --> 00:05:20,639
by filling in the template um I love

119
00:05:16,880 --> 00:05:24,919
this movie overall it was blank and then

120
00:05:20,639 --> 00:05:24,919
fill this uh continuation

121
00:05:25,840 --> 00:05:31,919
in a particular variety uh

122
00:05:30,000 --> 00:05:34,039
that we use very broadly nowadays

123
00:05:31,919 --> 00:05:36,240
because a lot of models are trained as

124
00:05:34,039 --> 00:05:38,240
chatbots um but actually even if they're

125
00:05:36,240 --> 00:05:41,199
not trained as chatbots this still works

126
00:05:38,240 --> 00:05:46,199
to some extent um is a chat

127
00:05:41,199 --> 00:05:49,919
prompt and so usually the way we we do

128
00:05:46,199 --> 00:05:53,240
this is we specify inputs in a format

129
00:05:49,919 --> 00:05:55,800
called the open AI messages format and

130
00:05:53,240 --> 00:05:58,199
uh this is this is what it looks like

131
00:05:55,800 --> 00:06:03,759
each we have a

132
00:05:58,199 --> 00:06:07,680
list of outputs each list is given a

133
00:06:03,759 --> 00:06:10,280
role and content and here so we have the

134
00:06:07,680 --> 00:06:12,479
role of system and the content is please

135
00:06:10,280 --> 00:06:15,319
classify movie reviews as positive or

136
00:06:12,479 --> 00:06:17,400
negative uh then we have the role user

137
00:06:15,319 --> 00:06:21,039
uh this movie is a

138
00:06:17,400 --> 00:06:24,919
banger um and then we have roles uh

139
00:06:21,039 --> 00:06:27,240
system message uh so is the roles we

140
00:06:24,919 --> 00:06:29,639
have the system and the system is a

141
00:06:27,240 --> 00:06:31,560
message provided to the system to

142
00:06:29,639 --> 00:06:33,560
influence Its Behavior it's to explain

143
00:06:31,560 --> 00:06:39,240
to it

144
00:06:33,560 --> 00:06:40,840
like how it should be working um and so

145
00:06:39,240 --> 00:06:43,199
you can see that this is explaining to

146
00:06:40,840 --> 00:06:46,400
the system how it should be working user

147
00:06:43,199 --> 00:06:48,680
is the message input by the user um and

148
00:06:46,400 --> 00:06:51,160
so this could be just a single message

149
00:06:48,680 --> 00:06:53,520
or if you have a multi-turn dialogue it

150
00:06:51,160 --> 00:06:55,080
can be like user and then assistant and

151
00:06:53,520 --> 00:06:56,680
then user and then assistant and then

152
00:06:55,080 --> 00:06:59,400
user and then assistant and that makes

153
00:06:56,680 --> 00:07:00,680
it clear that it's a multi-term dialogue

154
00:06:59,400 --> 00:07:02,800
so if you have a multi-term dialogue in

155
00:07:00,680 --> 00:07:06,319
chat GPT that's how they're feeding it

156
00:07:02,800 --> 00:07:06,319
in um into the

157
00:07:06,479 --> 00:07:12,440
system so what's happening behind the

158
00:07:08,840 --> 00:07:14,160
scenes with these chat prompts basically

159
00:07:12,440 --> 00:07:17,720
they're being converted into token

160
00:07:14,160 --> 00:07:19,680
strings and then fed into the model so

161
00:07:17,720 --> 00:07:21,800
despite the fact that this is fed in in

162
00:07:19,680 --> 00:07:23,560
this format and it makes you think that

163
00:07:21,800 --> 00:07:25,120
maybe something special is going on

164
00:07:23,560 --> 00:07:28,360
actually in most cases these are just

165
00:07:25,120 --> 00:07:30,199
being fed into the model uh as a prompt

166
00:07:28,360 --> 00:07:34,560
so these are just kind of special

167
00:07:30,199 --> 00:07:36,879
version of a uh of a template so here we

168
00:07:34,560 --> 00:07:40,560
have um this is what the Llama template

169
00:07:36,879 --> 00:07:43,319
looks like so basically you have um

170
00:07:40,560 --> 00:07:46,560
square bracket ins and then for the

171
00:07:43,319 --> 00:07:49,280
system message it's like um like angle

172
00:07:46,560 --> 00:07:51,240
bracket uh angle bracket sis uh close

173
00:07:49,280 --> 00:07:53,720
angle bracket close angle bracket and

174
00:07:51,240 --> 00:07:55,759
then the actual system message and then

175
00:07:53,720 --> 00:07:58,479
you have uh this closing out the system

176
00:07:55,759 --> 00:08:01,240
message this closing out the instruction

177
00:07:58,479 --> 00:08:04,120
then the user is surrounded by inst and

178
00:08:01,240 --> 00:08:06,599
then the assistant is just like a

179
00:08:04,120 --> 00:08:08,400
regular string so this is what the

180
00:08:06,599 --> 00:08:12,319
actual textual string that's fed into

181
00:08:08,400 --> 00:08:14,199
llama chat models is we can contrast

182
00:08:12,319 --> 00:08:19,440
that to some other models so alpaka

183
00:08:14,199 --> 00:08:22,400
looks like this um uh so we have like

184
00:08:19,440 --> 00:08:24,879
hash instruction colon and then the

185
00:08:22,400 --> 00:08:26,639
instruction for the user there there's

186
00:08:24,879 --> 00:08:28,879
no distinction between system and user

187
00:08:26,639 --> 00:08:31,960
so it's like hash instruction and then

188
00:08:28,879 --> 00:08:35,240
the user message and then hash response

189
00:08:31,960 --> 00:08:37,760
and then be assistant so it's not super

190
00:08:35,240 --> 00:08:39,640
important which one we use here um the

191
00:08:37,760 --> 00:08:41,919
important thing is that this matches

192
00:08:39,640 --> 00:08:44,039
with what uh the model is trained and

193
00:08:41,919 --> 00:08:46,640
I'll show you some example uh you know

194
00:08:44,039 --> 00:08:50,680
I'll talk about that in more detail

195
00:08:46,640 --> 00:08:52,880
later and we have a reference uh that I

196
00:08:50,680 --> 00:08:56,600
got this uh

197
00:08:52,880 --> 00:08:58,519
from and there's this toolkit that I um

198
00:08:56,600 --> 00:09:02,680
I rather like recently it's called light

199
00:08:58,519 --> 00:09:05,079
llm it makes it very easy to uh query

200
00:09:02,680 --> 00:09:07,240
different llms uh and kind of like

201
00:09:05,079 --> 00:09:09,320
unified things so basically you can

202
00:09:07,240 --> 00:09:11,800
query many different types of LMS like

203
00:09:09,320 --> 00:09:14,440
open AI or open source models or other

204
00:09:11,800 --> 00:09:17,079
things like that and what happens behind

205
00:09:14,440 --> 00:09:19,120
the scene is it basically takes um the

206
00:09:17,079 --> 00:09:20,839
open AI messages format and converts it

207
00:09:19,120 --> 00:09:22,880
into the appropriate prompt format for

208
00:09:20,839 --> 00:09:24,680
whatever model you're using or the

209
00:09:22,880 --> 00:09:27,120
appropriate API calls for whatever thing

210
00:09:24,680 --> 00:09:29,800
you're using but

211
00:09:27,120 --> 00:09:31,399
um this here basically

212
00:09:29,800 --> 00:09:33,800
um if you click through this link shows

213
00:09:31,399 --> 00:09:35,959
you okay this is what it looks like for

214
00:09:33,800 --> 00:09:37,880
alpaca um so you have the instruction

215
00:09:35,959 --> 00:09:40,920
instruction response this is what it

216
00:09:37,880 --> 00:09:44,880
looks like for llama 2 chat this is what

217
00:09:40,920 --> 00:09:48,480
it looks like for the oama um for AMA

218
00:09:44,880 --> 00:09:49,920
this is what it looks like for mistol

219
00:09:48,480 --> 00:09:52,160
and other things like that so you see

220
00:09:49,920 --> 00:09:53,440
all of these are very similar but

221
00:09:52,160 --> 00:09:55,000
they're like slightly different and

222
00:09:53,440 --> 00:09:58,120
getting these right is actually kind of

223
00:09:55,000 --> 00:10:01,120
important for the model doing a good

224
00:09:58,120 --> 00:10:01,120
job

225
00:10:03,640 --> 00:10:10,399
um any questions about

226
00:10:05,880 --> 00:10:15,360
this yeah like say you start PR with

227
00:10:10,399 --> 00:10:18,160
this um inut and then you started simar

228
00:10:15,360 --> 00:10:21,320
without

229
00:10:18,160 --> 00:10:24,640
model could you give an example yeah so

230
00:10:21,320 --> 00:10:28,040
say um my account is a great movie or

231
00:10:24,640 --> 00:10:31,040
this movie is great in front of I put

232
00:10:28,040 --> 00:10:31,040
UMR

233
00:10:34,279 --> 00:10:39,519
model

234
00:10:36,399 --> 00:10:42,440
so depend it depends a lot on the

235
00:10:39,519 --> 00:10:45,959
bottle the reason why this system

236
00:10:42,440 --> 00:10:48,720
message was input here in the first

237
00:10:45,959 --> 00:10:52,440
place was this wasn't originally a

238
00:10:48,720 --> 00:10:54,240
feature of open AI models uh open AI was

239
00:10:52,440 --> 00:10:56,440
the first place to introduce this which

240
00:10:54,240 --> 00:10:58,519
is why I I'm calling it open ey messages

241
00:10:56,440 --> 00:10:59,800
formul they didn't originally have

242
00:10:58,519 --> 00:11:02,360
something like this but they were having

243
00:10:59,800 --> 00:11:04,360
lots of trouble with um people trying to

244
00:11:02,360 --> 00:11:07,600
reveal the prompts that were given to

245
00:11:04,360 --> 00:11:09,680
systems uh like called like prompt

246
00:11:07,600 --> 00:11:12,040
injection attacks or like jailbreaking

247
00:11:09,680 --> 00:11:15,399
attacks or stff like that and so the

248
00:11:12,040 --> 00:11:17,079
models would basically reveal this

249
00:11:15,399 --> 00:11:19,600
prompt that was being used behind the

250
00:11:17,079 --> 00:11:22,760
scenes by whatever customer of open a

251
00:11:19,600 --> 00:11:26,120
was like deploying a system and so in

252
00:11:22,760 --> 00:11:29,120
order to fix this basically what open AI

253
00:11:26,120 --> 00:11:30,480
did I believe I believe like they're

254
00:11:29,120 --> 00:11:32,279
don't actually tell you exactly what

255
00:11:30,480 --> 00:11:36,040
they did ever but I'm assuming what they

256
00:11:32,279 --> 00:11:37,680
did is they trained uh their models so

257
00:11:36,040 --> 00:11:39,240
that the models would not output

258
00:11:37,680 --> 00:11:41,639
anything that's included in the system

259
00:11:39,240 --> 00:11:43,839
message so the system message is used to

260
00:11:41,639 --> 00:11:46,120
influence behavior but it like they're

261
00:11:43,839 --> 00:11:48,200
explicitly trained to not output things

262
00:11:46,120 --> 00:11:49,880
that are included in there and so if you

263
00:11:48,200 --> 00:11:53,360
put the

264
00:11:49,880 --> 00:11:56,200
actual if you put the actual thing that

265
00:11:53,360 --> 00:11:59,639
you wanted to evaluate within the system

266
00:11:56,200 --> 00:12:01,839
message it might still predict

267
00:11:59,639 --> 00:12:04,839
the sentiment correctly but it won't

268
00:12:01,839 --> 00:12:06,920
repeat the the stuff that was in system

269
00:12:04,839 --> 00:12:09,920
message

270
00:12:06,920 --> 00:12:09,920
B

271
00:12:14,160 --> 00:12:20,480
yeah after we give it the yeah yeah so

272
00:12:18,320 --> 00:12:23,040
the that's a great question so typically

273
00:12:20,480 --> 00:12:26,480
this is hand created so you you create

274
00:12:23,040 --> 00:12:29,680
something like this um I I have a a

275
00:12:26,480 --> 00:12:32,120
bracket X here but another way people

276
00:12:29,680 --> 00:12:33,800
typically specify this is you just have

277
00:12:32,120 --> 00:12:36,880
a

278
00:12:33,800 --> 00:12:41,199
big um you just have a big python string

279
00:12:36,880 --> 00:12:41,199
which is like um you know

280
00:12:42,040 --> 00:12:46,480
please um please

281
00:12:49,279 --> 00:12:55,440
specify and then you

282
00:12:52,440 --> 00:12:55,440
have

283
00:12:56,160 --> 00:13:02,240
um and then you substitute in uh like

284
00:12:59,880 --> 00:13:04,440
the input into this place here so you

285
00:13:02,240 --> 00:13:07,760
usually handw write it I'm going to

286
00:13:04,440 --> 00:13:07,760
talk excuse

287
00:13:07,800 --> 00:13:14,120
me and to end about some methods to

288
00:13:10,320 --> 00:13:16,120
learn these also but um I'd say like 90

289
00:13:14,120 --> 00:13:18,320
95% of the time people are just writing

290
00:13:16,120 --> 00:13:18,320
the

291
00:13:19,959 --> 00:13:24,560
man yep I would

292
00:13:25,920 --> 00:13:31,639
write

293
00:13:27,760 --> 00:13:31,639
and real input that

294
00:13:33,240 --> 00:13:38,040
I yeah so typically the template is

295
00:13:36,360 --> 00:13:39,800
written when you decide what system you

296
00:13:38,040 --> 00:13:41,839
want to create so you decide you want to

297
00:13:39,800 --> 00:13:44,519
create a sentiment analysis system so

298
00:13:41,839 --> 00:13:46,760
you create a template that either says

299
00:13:44,519 --> 00:13:48,079
like please classify the topic in the

300
00:13:46,760 --> 00:13:50,959
case of a model that was trained to

301
00:13:48,079 --> 00:13:52,240
follow instructions or if you have a

302
00:13:50,959 --> 00:13:54,240
base model that was not trained to

303
00:13:52,240 --> 00:13:58,079
follow instructions which is rare rare

304
00:13:54,240 --> 00:14:00,279
nowadays but gpd2 or La llama 2 without

305
00:13:58,079 --> 00:14:02,320
chat tuning is as an example of that

306
00:14:00,279 --> 00:14:05,600
then you would need to create a template

307
00:14:02,320 --> 00:14:10,040
that looks like this um where

308
00:14:05,600 --> 00:14:11,360
you put the model in a situation where

309
00:14:10,040 --> 00:14:13,839
the

310
00:14:11,360 --> 00:14:15,240
next word that follows up should be

311
00:14:13,839 --> 00:14:17,120
indicative of the answer to your

312
00:14:15,240 --> 00:14:20,120
question so like positive or negative or

313
00:14:17,120 --> 00:14:21,800
something like that so um but either way

314
00:14:20,120 --> 00:14:24,639
like usually you handw write this when

315
00:14:21,800 --> 00:14:27,199
you decide what task is you want to do

316
00:14:24,639 --> 00:14:29,000
then this input X this comes at test

317
00:14:27,199 --> 00:14:32,920
time this comes when you actually Dey

318
00:14:29,000 --> 00:14:34,240
your system um so this would be like an

319
00:14:32,920 --> 00:14:37,040
Amazon review that you wanted to

320
00:14:34,240 --> 00:14:37,040
classify using an

321
00:14:37,720 --> 00:14:42,720
image cool any other

322
00:14:40,519 --> 00:14:46,480
questions okay let's

323
00:14:42,720 --> 00:14:48,160
move um so basically this is what is

324
00:14:46,480 --> 00:14:49,920
happening behind the scenes I don't know

325
00:14:48,160 --> 00:14:53,040
what open AI format is because they

326
00:14:49,920 --> 00:14:54,639
won't tell us of course um but you know

327
00:14:53,040 --> 00:14:56,000
I'm assuming that that's similar to

328
00:14:54,639 --> 00:14:59,399
what's happening in

329
00:14:56,000 --> 00:15:01,959
op okay um so the next thing that we do

330
00:14:59,399 --> 00:15:05,360
is answer prediction so given uh The

331
00:15:01,959 --> 00:15:08,320
Prompt we predict the answer um and so

332
00:15:05,360 --> 00:15:11,880
using whatever algorithm we want to use

333
00:15:08,320 --> 00:15:14,880
uh we predict you know fantastic

334
00:15:11,880 --> 00:15:14,880
here

335
00:15:15,120 --> 00:15:21,639
um and actually it might not predict

336
00:15:19,959 --> 00:15:26,399
fantastic it might predict something

337
00:15:21,639 --> 00:15:28,120
else like overall it was um a really

338
00:15:26,399 --> 00:15:30,000
fantastic movie that I liked a lot or

339
00:15:28,120 --> 00:15:33,839
something like so it might also do

340
00:15:30,000 --> 00:15:36,880
something like that so based on that we

341
00:15:33,839 --> 00:15:39,600
want to select the actual output out of

342
00:15:36,880 --> 00:15:41,160
the generated uh outputs and I'm calling

343
00:15:39,600 --> 00:15:43,639
this uh

344
00:15:41,160 --> 00:15:45,959
postprocessing so for instance we might

345
00:15:43,639 --> 00:15:48,240
take the output as is so for something

346
00:15:45,959 --> 00:15:50,880
like just you interacting with chat

347
00:15:48,240 --> 00:15:53,360
jpt um or interacting with a chat model

348
00:15:50,880 --> 00:15:55,639
you might be looking at the text as is

349
00:15:53,360 --> 00:15:58,319
or it might be formatting the output for

350
00:15:55,639 --> 00:16:00,079
easy Vis visualization selecting only

351
00:15:58,319 --> 00:16:02,440
parts of the output that you want to use

352
00:16:00,079 --> 00:16:04,560
or mapping the output to other

353
00:16:02,440 --> 00:16:07,600
actions so to give an example of

354
00:16:04,560 --> 00:16:10,079
formatting this is a feature of uh chat

355
00:16:07,600 --> 00:16:13,440
GPT or Bard or any that you interact

356
00:16:10,079 --> 00:16:14,920
with but um I wrote please write a table

357
00:16:13,440 --> 00:16:18,759
with the last five presidents and their

358
00:16:14,920 --> 00:16:20,319
birth dates and chat GPT is happy to do

359
00:16:18,759 --> 00:16:22,000
this for me it says here is a table with

360
00:16:20,319 --> 00:16:24,920
the last five US presidents and their

361
00:16:22,000 --> 00:16:27,639
birth dates um Joe Biden Donald Trump

362
00:16:24,920 --> 00:16:31,720
Barack Obama George W wish Bill Clinton

363
00:16:27,639 --> 00:16:33,600
um but this is written in markdown um or

364
00:16:31,720 --> 00:16:35,079
I assume it's written in markdown so it

365
00:16:33,600 --> 00:16:37,880
basically makes this table and then

366
00:16:35,079 --> 00:16:39,319
renders it in an easy to view way so

367
00:16:37,880 --> 00:16:41,000
this is really important if you're

368
00:16:39,319 --> 00:16:42,440
building a user facing system because

369
00:16:41,000 --> 00:16:44,279
you want to be able to render these

370
00:16:42,440 --> 00:16:46,279
things but the only thing a large

371
00:16:44,279 --> 00:16:48,880
language model can output is text right

372
00:16:46,279 --> 00:16:50,279
it can output a string of tokens so uh

373
00:16:48,880 --> 00:16:54,000
this is a really good way to interact

374
00:16:50,279 --> 00:16:55,759
with it um I I followed by saying output

375
00:16:54,000 --> 00:16:58,720
that in Json format so it says here's

376
00:16:55,759 --> 00:17:00,360
the information in Json format and

377
00:16:58,720 --> 00:17:02,000
instead of just giving me a big Json

378
00:17:00,360 --> 00:17:04,199
string it gives me syntax highlighting

379
00:17:02,000 --> 00:17:06,880
and all the other stuff like this um

380
00:17:04,199 --> 00:17:09,760
presumably what it's doing here is it's

381
00:17:06,880 --> 00:17:12,839
outputting um like a triple hash or

382
00:17:09,760 --> 00:17:15,160
something like this um the reason why I

383
00:17:12,839 --> 00:17:17,600
know that is because

384
00:17:15,160 --> 00:17:21,079
like seems to be making a mistake down

385
00:17:17,600 --> 00:17:23,280
here for some reason um like uh

386
00:17:21,079 --> 00:17:25,079
outputting a weird Le formatted thing at

387
00:17:23,280 --> 00:17:26,160
that and so even chat GPT makes mistakes

388
00:17:25,079 --> 00:17:30,320
some of the

389
00:17:26,160 --> 00:17:32,400
time um

390
00:17:30,320 --> 00:17:33,960
cool um another thing that you might

391
00:17:32,400 --> 00:17:35,520
want to do is especially if you're not

392
00:17:33,960 --> 00:17:37,360
using it in like a a directly

393
00:17:35,520 --> 00:17:40,200
user-facing application but you want to

394
00:17:37,360 --> 00:17:41,840
use it to extract some information or

395
00:17:40,200 --> 00:17:45,440
make some classification decision or

396
00:17:41,840 --> 00:17:47,280
something like that um you often select

397
00:17:45,440 --> 00:17:49,880
information that's indicative of the

398
00:17:47,280 --> 00:17:52,360
answer and so I love this movie overall

399
00:17:49,880 --> 00:17:53,960
it was a movie that was simply fantastic

400
00:17:52,360 --> 00:17:56,600
um you can do things like extract

401
00:17:53,960 --> 00:17:59,440
keywords like fantastic and use that to

402
00:17:56,600 --> 00:18:01,360
indicate positive sentiment

403
00:17:59,440 --> 00:18:04,080
there's various methods for doing this

404
00:18:01,360 --> 00:18:05,919
and these are also used in the

405
00:18:04,080 --> 00:18:08,679
benchmarks that are used to evaluate

406
00:18:05,919 --> 00:18:09,799
language models so it's you know like

407
00:18:08,679 --> 00:18:11,039
even if you're not building an

408
00:18:09,799 --> 00:18:12,679
application directly but you're just

409
00:18:11,039 --> 00:18:14,120
trying to do well in this class and get

410
00:18:12,679 --> 00:18:15,679
like a high score on a leaderboard or

411
00:18:14,120 --> 00:18:20,320
something it's still useful to know

412
00:18:15,679 --> 00:18:22,159
about these things so um for things like

413
00:18:20,320 --> 00:18:24,039
classification um you can identify

414
00:18:22,159 --> 00:18:27,159
keywords like fantastic that might be

415
00:18:24,039 --> 00:18:29,120
indicative of the class another thing

416
00:18:27,159 --> 00:18:31,559
that's uh pretty common is for

417
00:18:29,120 --> 00:18:34,480
regression or numerical problems you

418
00:18:31,559 --> 00:18:37,440
identify numbers and pull out the

419
00:18:34,480 --> 00:18:40,400
numbers and use those numbers as the

420
00:18:37,440 --> 00:18:42,360
answer um for code uh you can pull out

421
00:18:40,400 --> 00:18:45,080
code Snippets and triple back ticks and

422
00:18:42,360 --> 00:18:46,960
then execute the code for example so all

423
00:18:45,080 --> 00:18:48,600
of these things are basically heuristic

424
00:18:46,960 --> 00:18:50,159
methods but they can be used to pull out

425
00:18:48,600 --> 00:18:53,440
the actual answer that you want from the

426
00:18:50,159 --> 00:18:53,440
text that's generated di

427
00:18:54,480 --> 00:19:00,320
know cool uh any questions about that

428
00:19:02,280 --> 00:19:07,880
the final thing is output mapping um

429
00:19:04,640 --> 00:19:11,120
given an answer uh map it into a class

430
00:19:07,880 --> 00:19:13,360
label or a continuous value and so this

431
00:19:11,120 --> 00:19:16,000
is doing something like taking fantastic

432
00:19:13,360 --> 00:19:18,480
and mapping it into the class

433
00:19:16,000 --> 00:19:21,000
positive uh and so you know if we want

434
00:19:18,480 --> 00:19:23,000
to extract fi one to five star ratings

435
00:19:21,000 --> 00:19:25,559
from reviews this is something you would

436
00:19:23,000 --> 00:19:29,360
need to do and very often it's like a

437
00:19:25,559 --> 00:19:33,880
one to um one class to

438
00:19:29,360 --> 00:19:35,720
many um many word mapping and uh by

439
00:19:33,880 --> 00:19:37,400
doing this you can basically get a more

440
00:19:35,720 --> 00:19:38,720
robust mapping onto the number that you

441
00:19:37,400 --> 00:19:42,400
actually

442
00:19:38,720 --> 00:19:42,400
want I actually

443
00:19:42,720 --> 00:19:48,919
coincidentally on uh on Twitter saw a

444
00:19:45,280 --> 00:19:48,919
really good example of this like a week

445
00:19:55,880 --> 00:20:00,520
ago and yeah I don't know if I'm going

446
00:19:59,120 --> 00:20:05,440
to be able to find it in a reasonable

447
00:20:00,520 --> 00:20:08,520
time frame but basically um there was

448
00:20:05,440 --> 00:20:11,080
a person who was using gp4 to create a

449
00:20:08,520 --> 00:20:14,120
model uh to like reward open source

450
00:20:11,080 --> 00:20:15,880
models for good and bad you know

451
00:20:14,120 --> 00:20:18,320
responses

452
00:20:15,880 --> 00:20:20,799
and they started out with giving it a

453
00:20:18,320 --> 00:20:24,480
one to five star rating and then they

454
00:20:20,799 --> 00:20:28,360
switched it into very good good okay bad

455
00:20:24,480 --> 00:20:31,280
very bad and then um then asked to

456
00:20:28,360 --> 00:20:34,520
generate you know those like very good

457
00:20:31,280 --> 00:20:37,039
good bad okay bad very bad instead of

458
00:20:34,520 --> 00:20:40,360
one to five and that worked a lot better

459
00:20:37,039 --> 00:20:43,480
like the GPT model was a lot more uh

460
00:20:40,360 --> 00:20:46,039
like likely to get the answer correct um

461
00:20:43,480 --> 00:20:48,880
than it was if you gave a one to five

462
00:20:46,039 --> 00:20:50,799
star rating so this is something you

463
00:20:48,880 --> 00:20:54,280
should think about pretty seriously and

464
00:20:50,799 --> 00:20:57,440
the way you can think about it is How

465
00:20:54,280 --> 00:20:59,679
likely was this data to appear in a

466
00:20:57,440 --> 00:21:02,520
large Corp of data on the

467
00:20:59,679 --> 00:21:04,760
internet and it might be like a lot less

468
00:21:02,520 --> 00:21:08,679
likely that it's like how good is this

469
00:21:04,760 --> 00:21:11,400
movie five then how good is this movie

470
00:21:08,679 --> 00:21:13,960
really good like just think of like the

471
00:21:11,400 --> 00:21:16,200
occurrence probability and you can even

472
00:21:13,960 --> 00:21:18,600
um like mine this data from the the web

473
00:21:16,200 --> 00:21:21,320
if you want to to try to find out the

474
00:21:18,600 --> 00:21:24,520
best you know

475
00:21:21,320 --> 00:21:30,039
like the best things

476
00:21:24,520 --> 00:21:30,039
there cool um any questions about this

477
00:21:35,360 --> 00:21:39,480
yeah how is

478
00:21:37,720 --> 00:21:43,039
it

479
00:21:39,480 --> 00:21:45,919
learning so the model the model is

480
00:21:43,039 --> 00:21:47,600
predicting txt and like accurately it's

481
00:21:45,919 --> 00:21:50,200
not even predicting the word fantastic

482
00:21:47,600 --> 00:21:54,480
it's predicting the token ID like

483
00:21:50,200 --> 00:21:57,600
73521 or something like that um but you

484
00:21:54,480 --> 00:21:58,679
know if it has seen that token ID more

485
00:21:57,600 --> 00:22:00,840
frequent

486
00:21:58,679 --> 00:22:04,240
after reviews than it has seen the token

487
00:22:00,840 --> 00:22:06,000
ID for the number one or the number five

488
00:22:04,240 --> 00:22:07,520
then it's more likely to predict that

489
00:22:06,000 --> 00:22:10,279
accurately right it's more likely to

490
00:22:07,520 --> 00:22:11,880
predict fantastic than it is to predict

491
00:22:10,279 --> 00:22:14,679
five star or something like that just

492
00:22:11,880 --> 00:22:16,720
because fantastic is more frequent and

493
00:22:14,679 --> 00:22:18,880
so because of that if you think about

494
00:22:16,720 --> 00:22:22,120
like what has it seen in all of the data

495
00:22:18,880 --> 00:22:24,240
on the internet and like model your um

496
00:22:22,120 --> 00:22:26,960
model your answers here appropriately

497
00:22:24,240 --> 00:22:28,520
then that can give you

498
00:22:26,960 --> 00:22:30,320
betters

499
00:22:28,520 --> 00:22:32,120
this is a very important rule of thumb

500
00:22:30,320 --> 00:22:33,400
like don't try to make a language model

501
00:22:32,120 --> 00:22:35,039
do something it's never seen in the

502
00:22:33,400 --> 00:22:38,200
pre-training data and it will make your

503
00:22:35,039 --> 00:22:40,240
life a lot easier so um you can think

504
00:22:38,200 --> 00:22:41,880
that going forward

505
00:22:40,240 --> 00:22:44,679
to

506
00:22:41,880 --> 00:22:48,559
cool so next I want to move into fat

507
00:22:44,679 --> 00:22:49,679
prompting or in context learning um so

508
00:22:48,559 --> 00:22:52,159
fat

509
00:22:49,679 --> 00:22:54,440
prompting basically what we do is we

510
00:22:52,159 --> 00:22:55,799
provide a few examples of the task

511
00:22:54,440 --> 00:22:58,440
together with the

512
00:22:55,799 --> 00:23:00,080
instruction and the way this work works

513
00:22:58,440 --> 00:23:02,360
is you write an instruction like please

514
00:23:00,080 --> 00:23:05,919
classify movie reviews as positive or

515
00:23:02,360 --> 00:23:08,120
negative and add like input uh I really

516
00:23:05,919 --> 00:23:10,320
don't like this movie output negative uh

517
00:23:08,120 --> 00:23:12,480
input this movie is great output

518
00:23:10,320 --> 00:23:16,640
positive

519
00:23:12,480 --> 00:23:18,880
and this is um pretty effective the

520
00:23:16,640 --> 00:23:21,799
thing it's most effective for are

521
00:23:18,880 --> 00:23:24,400
twofold it's most effective for making

522
00:23:21,799 --> 00:23:26,360
sure that you get the formatting right

523
00:23:24,400 --> 00:23:27,640
uh because if you have a few examples

524
00:23:26,360 --> 00:23:28,679
the model will tend to follow those

525
00:23:27,640 --> 00:23:30,840
examples

526
00:23:28,679 --> 00:23:34,440
with respect to formatting especially if

527
00:23:30,840 --> 00:23:37,320
we're talking about like gp4 models um

528
00:23:34,440 --> 00:23:40,400
or strong GPT models it's also effective

529
00:23:37,320 --> 00:23:42,400
if you're using weaker models so like

530
00:23:40,400 --> 00:23:44,720
stronger models like gp4 tend to be

531
00:23:42,400 --> 00:23:46,720
pretty good at following instructions so

532
00:23:44,720 --> 00:23:49,520
if you say

533
00:23:46,720 --> 00:23:51,640
um please classify movie reviews as

534
00:23:49,520 --> 00:23:54,000
positive or negative it will be more

535
00:23:51,640 --> 00:23:56,279
likely to just output positive or

536
00:23:54,000 --> 00:23:58,760
negative um but if you have weaker

537
00:23:56,279 --> 00:24:01,720
models it might say I really don't like

538
00:23:58,760 --> 00:24:03,559
this movie output uh I think I think

539
00:24:01,720 --> 00:24:05,640
this is probably negative or something

540
00:24:03,559 --> 00:24:07,240
like that it will you know it might not

541
00:24:05,640 --> 00:24:10,080
follow the instructions as well and it's

542
00:24:07,240 --> 00:24:14,240
more effective to provide as in context

543
00:24:10,080 --> 00:24:17,600
examples um so so this is a one uh

544
00:24:14,240 --> 00:24:19,480
one uh thing to remember one thing I

545
00:24:17,600 --> 00:24:22,120
should mention also is when I say F shot

546
00:24:19,480 --> 00:24:25,720
prompting and in context learning these

547
00:24:22,120 --> 00:24:27,880
are basically the same thing uh they

548
00:24:25,720 --> 00:24:29,720
basically refer to the same concept but

549
00:24:27,880 --> 00:24:31,919
just from slightly different

550
00:24:29,720 --> 00:24:34,799
examples uh from sorry slightly

551
00:24:31,919 --> 00:24:36,919
different angles PE shot is in contrast

552
00:24:34,799 --> 00:24:39,320
to zero shot so zero shot means you're

553
00:24:36,919 --> 00:24:43,039
providing no examples so zero shot

554
00:24:39,320 --> 00:24:45,720
prompting you would have none uh few

555
00:24:43,039 --> 00:24:47,240
shot you have several examples in

556
00:24:45,720 --> 00:24:49,679
context learning means that you're

557
00:24:47,240 --> 00:24:51,640
learning how to do a task but instead of

558
00:24:49,679 --> 00:24:54,320
providing the model with fine-tuning

559
00:24:51,640 --> 00:24:56,679
data you're providing the examples in

560
00:24:54,320 --> 00:24:58,080
the language models context so they both

561
00:24:56,679 --> 00:25:00,919
basically mean the same thing but

562
00:24:58,080 --> 00:25:03,159
they're they're just contrasting to like

563
00:25:00,919 --> 00:25:06,559
either a zero shot or fine tuning which

564
00:25:03,159 --> 00:25:06,559
is why the terminology is

565
00:25:06,880 --> 00:25:13,520
different so they usering interface

566
00:25:11,320 --> 00:25:16,080
and for the

567
00:25:13,520 --> 00:25:17,760
rendering uh yes you can definitely do F

568
00:25:16,080 --> 00:25:20,039
shot prompting I'm actually going to

569
00:25:17,760 --> 00:25:23,440
talk exactly about exactly how you do

570
00:25:20,039 --> 00:25:26,320
this in like an open AI model um here

571
00:25:23,440 --> 00:25:28,240
which is for open AI models there's a

572
00:25:26,320 --> 00:25:31,320
couple ways that you could do this one

573
00:25:28,240 --> 00:25:33,640
way you could do this is you could um

574
00:25:31,320 --> 00:25:36,279
you could have the role be user and the

575
00:25:33,640 --> 00:25:39,279
role be assistant and just add like

576
00:25:36,279 --> 00:25:41,159
additional conversational history into

577
00:25:39,279 --> 00:25:43,159
the the messages that you're sending to

578
00:25:41,159 --> 00:25:46,240
the language model but actually the

579
00:25:43,159 --> 00:25:49,120
recommended way of doing this um which

580
00:25:46,240 --> 00:25:51,880
is in the openi cookbook uh which is in

581
00:25:49,120 --> 00:25:53,919
the reference is that you send this as a

582
00:25:51,880 --> 00:25:58,200
system message but you provide this like

583
00:25:53,919 --> 00:26:00,840
additional name variable here um with

584
00:25:58,200 --> 00:26:02,840
example user and example assistant the

585
00:26:00,840 --> 00:26:06,200
main reason why you do this is just

586
00:26:02,840 --> 00:26:08,080
because if you don't um if you send it

587
00:26:06,200 --> 00:26:10,600
in as the like user and assistant the

588
00:26:08,080 --> 00:26:12,799
model might refer back to the few shot

589
00:26:10,600 --> 00:26:14,320
examples as something that happened

590
00:26:12,799 --> 00:26:15,760
previously in the conversation whereas

591
00:26:14,320 --> 00:26:18,200
if you send it in the system message

592
00:26:15,760 --> 00:26:19,799
it's guaranteed to not do that so I

593
00:26:18,200 --> 00:26:23,600
think it's like less of an accuracy

594
00:26:19,799 --> 00:26:26,360
thing it's more of a like it's more of a

595
00:26:23,600 --> 00:26:29,120
privacy prompt privacy thing uh than

596
00:26:26,360 --> 00:26:30,880
anything else so this is a recommended

597
00:26:29,120 --> 00:26:33,159
way of doing this on the other hand if

598
00:26:30,880 --> 00:26:34,600
you're using like an open source model

599
00:26:33,159 --> 00:26:36,600
uh you need to be careful because this

600
00:26:34,600 --> 00:26:38,279
name might not even be included in the

601
00:26:36,600 --> 00:26:40,080
prompt template like for example in the

602
00:26:38,279 --> 00:26:41,840
light llm prompt templates that I was

603
00:26:40,080 --> 00:26:44,080
sending in this is not even included at

604
00:26:41,840 --> 00:26:46,480
all so you might just get a weird system

605
00:26:44,080 --> 00:26:49,720
message that uh is poorly fored so you

606
00:26:46,480 --> 00:26:53,600
need to be a little bit conscious

607
00:26:49,720 --> 00:26:55,799
this um cool any questions here does

608
00:26:53,600 --> 00:26:58,880
that answer the

609
00:26:55,799 --> 00:27:02,279
question okay

610
00:26:58,880 --> 00:27:05,000
um so one one thing to be aware of is

611
00:27:02,279 --> 00:27:07,039
llms are sensitive to small changes and

612
00:27:05,000 --> 00:27:12,080
in context examples that you provide to

613
00:27:07,039 --> 00:27:14,600
them so uh previous work has examined

614
00:27:12,080 --> 00:27:19,399
this from a number of angles there's a

615
00:27:14,600 --> 00:27:22,679
paper by Luol and they examine the

616
00:27:19,399 --> 00:27:25,000
sensitivity to example ordering so like

617
00:27:22,679 --> 00:27:28,399
if you take the same examples and you

618
00:27:25,000 --> 00:27:30,840
just order them in different orders um

619
00:27:28,399 --> 00:27:32,679
you can actually get very wildly

620
00:27:30,840 --> 00:27:35,600
different

621
00:27:32,679 --> 00:27:37,520
results um and this is especially true

622
00:27:35,600 --> 00:27:40,320
for smaller models so the smaller models

623
00:27:37,520 --> 00:27:42,720
here are like the gpt2 models the larger

624
00:27:40,320 --> 00:27:47,440
models here are like the GPT the larger

625
00:27:42,720 --> 00:27:47,440
model here is GPT 3.5 uh I

626
00:27:48,399 --> 00:27:54,120
believe other things that people have

627
00:27:50,559 --> 00:27:56,760
looked at are label balance so um how

628
00:27:54,120 --> 00:27:58,559
important is it for the labels to be

629
00:27:56,760 --> 00:28:01,440
balanced

630
00:27:58,559 --> 00:28:02,799
um and if you're doing sentiment

631
00:28:01,440 --> 00:28:05,240
classification for example you might

632
00:28:02,799 --> 00:28:07,519
have only positive examples or only

633
00:28:05,240 --> 00:28:10,000
negative examples and if you have only

634
00:28:07,519 --> 00:28:13,279
positive or negative examples this can

635
00:28:10,000 --> 00:28:15,559
uh help or hurt your accuracy uh for

636
00:28:13,279 --> 00:28:17,200
example on this Amazon review data set

637
00:28:15,559 --> 00:28:18,679
most of the reviews are positive so you

638
00:28:17,200 --> 00:28:20,840
actually do better by having lots of

639
00:28:18,679 --> 00:28:23,640
positive examples in your in context

640
00:28:20,840 --> 00:28:26,600
examples on the other hand for sst2 this

641
00:28:23,640 --> 00:28:29,159
is label balanced so having only

642
00:28:26,600 --> 00:28:31,799
positive or negative is worse on average

643
00:28:29,159 --> 00:28:34,279
than having three positive and one

644
00:28:31,799 --> 00:28:36,679
negative another thing is label coverage

645
00:28:34,279 --> 00:28:38,679
so if we're talking about multi class

646
00:28:36,679 --> 00:28:41,120
classification um

647
00:28:38,679 --> 00:28:42,919
having good coverage of all of the

648
00:28:41,120 --> 00:28:45,919
classes that you want to include in your

649
00:28:42,919 --> 00:28:49,120
multiclass classification is important

650
00:28:45,919 --> 00:28:51,720
um to some extent but if you have uh

651
00:28:49,120 --> 00:28:53,440
more uh you can also confuse some model

652
00:28:51,720 --> 00:28:55,840
especially if they're minority labels so

653
00:28:53,440 --> 00:28:57,799
if you have a whole bunch of like random

654
00:28:55,840 --> 00:28:59,080
minority labels and that can cause so

655
00:28:57,799 --> 00:29:01,399
this is something important to think

656
00:28:59,080 --> 00:29:04,640
about if you're planning on solving kind

657
00:29:01,399 --> 00:29:08,640
of like classification tests um I I've

658
00:29:04,640 --> 00:29:11,000
also had my own experience with uh using

659
00:29:08,640 --> 00:29:13,159
GPT for evaluation for machine

660
00:29:11,000 --> 00:29:14,760
translation and when we use GPT for

661
00:29:13,159 --> 00:29:18,559
evaluation for machine translation it

662
00:29:14,760 --> 00:29:20,799
was very important to add um like high

663
00:29:18,559 --> 00:29:22,760
uh high scoring values low score high

664
00:29:20,799 --> 00:29:26,320
scoring outputs low scoring outputs some

665
00:29:22,760 --> 00:29:27,840
in the middle um and so it's also the

666
00:29:26,320 --> 00:29:30,760
case for regression

667
00:29:27,840 --> 00:29:30,760
uh problems as

668
00:29:32,600 --> 00:29:37,320
well cool um any questions

669
00:29:38,159 --> 00:29:45,000
here um however this is not super

670
00:29:42,240 --> 00:29:46,600
predictable um so there's not like any

671
00:29:45,000 --> 00:29:48,399
rule of thumb that tells you like this

672
00:29:46,600 --> 00:29:49,720
is or as far as I know there's not any

673
00:29:48,399 --> 00:29:51,640
rule of thumb that tells you this is the

674
00:29:49,720 --> 00:29:54,000
way you should construct in context

675
00:29:51,640 --> 00:29:55,880
examples uh there are lots of papers

676
00:29:54,000 --> 00:29:57,799
that say they have methods that work

677
00:29:55,880 --> 00:30:01,000
better but I don't know if there's any

678
00:29:57,799 --> 00:30:02,559
like gold standard IND indry practice

679
00:30:01,000 --> 00:30:05,799
for doing something like this at the

680
00:30:02,559 --> 00:30:07,799
moment so just to give an example uh

681
00:30:05,799 --> 00:30:10,399
this paper it's a really nice paper

682
00:30:07,799 --> 00:30:13,440
examining why uh in context Learning

683
00:30:10,399 --> 00:30:17,279
Works one thing one interesting finding

684
00:30:13,440 --> 00:30:19,760
that they have is they output they take

685
00:30:17,279 --> 00:30:22,720
in context examples but they randomize

686
00:30:19,760 --> 00:30:27,320
the labels they make the labels wrong

687
00:30:22,720 --> 00:30:29,519
some of the time so even with completely

688
00:30:27,320 --> 00:30:32,120
wrong labels even with labels that are

689
00:30:29,519 --> 00:30:34,399
correct 0% of the time you still get

690
00:30:32,120 --> 00:30:37,360
much much better accuracy than if you

691
00:30:34,399 --> 00:30:39,440
use no Inc context examples and why is

692
00:30:37,360 --> 00:30:41,640
this probably you know it's getting the

693
00:30:39,440 --> 00:30:44,600
model formatting correct it's getting

694
00:30:41,640 --> 00:30:47,679
like the names of the labels correct

695
00:30:44,600 --> 00:30:49,039
even if it's not uh accurate so it seems

696
00:30:47,679 --> 00:30:50,519
like it's not really using these for

697
00:30:49,039 --> 00:30:52,640
training data it's using them more just

698
00:30:50,519 --> 00:30:56,240
to know the formatting

699
00:30:52,640 --> 00:30:59,399
appropriate like

700
00:30:56,240 --> 00:31:01,399
you so you already

701
00:30:59,399 --> 00:31:03,760
have

702
00:31:01,399 --> 00:31:08,840
right how is it

703
00:31:03,760 --> 00:31:11,240
Ma like is it just y one y i gu I'm just

704
00:31:08,840 --> 00:31:15,000
ask how you would inter

705
00:31:11,240 --> 00:31:16,480
that so this is you're not training the

706
00:31:15,000 --> 00:31:17,880
model at the moment we're going to talk

707
00:31:16,480 --> 00:31:19,360
about that next class but right now

708
00:31:17,880 --> 00:31:21,279
you're taking a model that has already

709
00:31:19,360 --> 00:31:22,840
been trained and you're providing it

710
00:31:21,279 --> 00:31:25,519
with a few examples and then you're

711
00:31:22,840 --> 00:31:28,679
asking it to fill in um the following

712
00:31:25,519 --> 00:31:30,880
examples just examples

713
00:31:28,679 --> 00:31:32,960
yes

714
00:31:30,880 --> 00:31:34,679
exactly and it's pretty amazing that

715
00:31:32,960 --> 00:31:36,440
that works in the first place especially

716
00:31:34,679 --> 00:31:39,840
with a model that hasn't been explicitly

717
00:31:36,440 --> 00:31:41,200
trained that way but um there's a a fair

718
00:31:39,840 --> 00:31:42,320
amount of research that I think we're

719
00:31:41,200 --> 00:31:43,960
probably going to be talking about in

720
00:31:42,320 --> 00:31:47,000
the interpretability class about why

721
00:31:43,960 --> 00:31:49,600
this happens but um

722
00:31:47,000 --> 00:31:51,279
basically my my interpretation for why

723
00:31:49,600 --> 00:31:53,679
this happens is because there's so much

724
00:31:51,279 --> 00:31:56,000
repetitive stuff on the internet right

725
00:31:53,679 --> 00:31:58,240
there's a bunch of examples of math

726
00:31:56,000 --> 00:32:00,399
problems which is like

727
00:31:58,240 --> 00:32:02,279
question one and then the math problem

728
00:32:00,399 --> 00:32:04,320
and then the answer question two math

729
00:32:02,279 --> 00:32:06,440
problem and then the answer so in order

730
00:32:04,320 --> 00:32:08,320
to model the text on the internet it

731
00:32:06,440 --> 00:32:12,120
needs to learn how to be able to do

732
00:32:08,320 --> 00:32:15,399
these things but so um

733
00:32:12,120 --> 00:32:17,760
cool the second thing is uh more

734
00:32:15,399 --> 00:32:20,000
demonstrations can sometimes hurt

735
00:32:17,760 --> 00:32:22,120
accuracy so this is like binary

736
00:32:20,000 --> 00:32:25,080
classification versus multiple choice

737
00:32:22,120 --> 00:32:27,440
question answering um and actually with

738
00:32:25,080 --> 00:32:30,919
binary classification the model ends up

739
00:32:27,440 --> 00:32:33,159
getting worse um with uh more examples

740
00:32:30,919 --> 00:32:36,799
probably just because the longer context

741
00:32:33,159 --> 00:32:39,320
uh you know confuses the model or moves

742
00:32:36,799 --> 00:32:41,320
the instructions that are provided to

743
00:32:39,320 --> 00:32:44,279
the model farther away in the context so

744
00:32:41,320 --> 00:32:48,120
it starts forgetting them so

745
00:32:44,279 --> 00:32:50,240
um basically what I want to say is uh

746
00:32:48,120 --> 00:32:51,760
you know this is more of an art than a

747
00:32:50,240 --> 00:32:53,279
science you might not get entirely

748
00:32:51,760 --> 00:32:55,840
predictable results but don't worry it's

749
00:32:53,279 --> 00:32:59,320
not just

750
00:32:55,840 --> 00:32:59,320
you cool cool

751
00:33:09,200 --> 00:33:15,320
yeah it can't so the question is if the

752
00:33:12,639 --> 00:33:17,039
in context examples reflect the data

753
00:33:15,320 --> 00:33:18,919
distribution well would that boost the

754
00:33:17,039 --> 00:33:24,240
accuracy I think the answer is probably

755
00:33:18,919 --> 00:33:26,039
yes yeah um I don't know if that it's

756
00:33:24,240 --> 00:33:27,679
that clear because like what I would

757
00:33:26,039 --> 00:33:29,919
expect

758
00:33:27,679 --> 00:33:33,559
is better

759
00:33:29,919 --> 00:33:37,240
coverage is probably more

760
00:33:33,559 --> 00:33:39,760
important than better representativeness

761
00:33:37,240 --> 00:33:41,960
so like even if you have some minority

762
00:33:39,760 --> 00:33:43,639
labels um it's probably better for the

763
00:33:41,960 --> 00:33:44,880
model to know what those minority labels

764
00:33:43,639 --> 00:33:47,279
look like and that's going to be

765
00:33:44,880 --> 00:33:49,120
especially true for like stronger models

766
00:33:47,279 --> 00:33:50,679
um I think

767
00:33:49,120 --> 00:33:54,320
so

768
00:33:50,679 --> 00:33:56,440
cool okay so uh next I want to talk

769
00:33:54,320 --> 00:33:59,000
about Chain of Thought prompting um so

770
00:33:56,440 --> 00:34:01,320
Chain of Thought prompting is a very

771
00:33:59,000 --> 00:34:04,080
popular way of prompting

772
00:34:01,320 --> 00:34:06,080
models and the way it works is you get

773
00:34:04,080 --> 00:34:07,839
the model to explain its reasoning

774
00:34:06,080 --> 00:34:12,679
before making an

775
00:34:07,839 --> 00:34:14,520
answer um and so sorry this example is a

776
00:34:12,679 --> 00:34:18,879
little bit small but like the standard

777
00:34:14,520 --> 00:34:20,480
prompting method is uh like Roger has

778
00:34:18,879 --> 00:34:22,000
five tennis balls he buys two more cans

779
00:34:20,480 --> 00:34:23,480
of tennis balls each can has three

780
00:34:22,000 --> 00:34:28,200
tennis balls how many tennis balls does

781
00:34:23,480 --> 00:34:29,359
he have now um the answer is 11 and so

782
00:34:28,200 --> 00:34:32,119
um this

783
00:34:29,359 --> 00:34:34,320
is an in context example and then you

784
00:34:32,119 --> 00:34:37,240
have your input which has a different

785
00:34:34,320 --> 00:34:39,000
problem uh the cafeteria has 23 apples

786
00:34:37,240 --> 00:34:40,639
if they Ed 20 to make lunch and bought

787
00:34:39,000 --> 00:34:41,800
six more how many apples do they have

788
00:34:40,639 --> 00:34:46,720
the answer is

789
00:34:41,800 --> 00:34:49,000
27 um and so this is wrong so what Chain

790
00:34:46,720 --> 00:34:52,000
of Thought prompting does is instead of

791
00:34:49,000 --> 00:34:54,960
just giving the answer it gives you an

792
00:34:52,000 --> 00:34:57,079
additional reasoning chain uh that says

793
00:34:54,960 --> 00:34:59,680
R started with five balls two cans of of

794
00:34:57,079 --> 00:35:01,800
three tennis balls uh each of six tennis

795
00:34:59,680 --> 00:35:04,520
balls 5 plus 6 equals 11 the answer is

796
00:35:01,800 --> 00:35:06,280
11 and so then when you feed this in

797
00:35:04,520 --> 00:35:08,000
basically the model will generate a

798
00:35:06,280 --> 00:35:10,240
similar reasoning chain and then it's

799
00:35:08,000 --> 00:35:13,400
more likely to get the answer correct

800
00:35:10,240 --> 00:35:15,720
and this very robustly works

801
00:35:13,400 --> 00:35:19,440
for many

802
00:35:15,720 --> 00:35:21,440
different problems where a reasoning

803
00:35:19,440 --> 00:35:23,520
chain is

804
00:35:21,440 --> 00:35:27,839
necessary and if you think about the

805
00:35:23,520 --> 00:35:30,359
reason why this uh why this works I

806
00:35:27,839 --> 00:35:33,040
think there's basically two reasons why

807
00:35:30,359 --> 00:35:34,440
um the first reason is I I only wrote

808
00:35:33,040 --> 00:35:36,560
one on the thing here but the first

809
00:35:34,440 --> 00:35:38,760
reason is it allows the model to

810
00:35:36,560 --> 00:35:41,359
decompose harder problems into simpler

811
00:35:38,760 --> 00:35:45,119
problems and simpler problems are easier

812
00:35:41,359 --> 00:35:47,560
right so um instead

813
00:35:45,119 --> 00:35:51,319
of immediately trying to solve the whole

814
00:35:47,560 --> 00:35:53,800
problem in a single go it will first

815
00:35:51,319 --> 00:35:56,520
solve the problem of like what how many

816
00:35:53,800 --> 00:35:58,920
are left after you use buy and so it

817
00:35:56,520 --> 00:36:00,240
gets three and so now it has this three

818
00:35:58,920 --> 00:36:02,480
here so now it can solve the next

819
00:36:00,240 --> 00:36:05,160
problem of adding six that's equal to 9

820
00:36:02,480 --> 00:36:07,880
so it's solving simpler sub problems

821
00:36:05,160 --> 00:36:11,440
than it is and uh compared to harder

822
00:36:07,880 --> 00:36:13,920
ones another reason why is it allows for

823
00:36:11,440 --> 00:36:17,319
adaptive computation time so if you

824
00:36:13,920 --> 00:36:17,319
think about like a Transformer

825
00:36:19,000 --> 00:36:23,119
model um if you think about a

826
00:36:21,280 --> 00:36:25,560
Transformer model a Transformer model

827
00:36:23,119 --> 00:36:27,200
has fixed computation time for

828
00:36:25,560 --> 00:36:29,920
predicting each token right a fixed

829
00:36:27,200 --> 00:36:31,560
number of layers it um and based on that

830
00:36:29,920 --> 00:36:33,839
fixed number of layers it passes all the

831
00:36:31,560 --> 00:36:36,520
information through and makes a

832
00:36:33,839 --> 00:36:38,200
prediction and some problems are harder

833
00:36:36,520 --> 00:36:39,599
than others right so it would be very

834
00:36:38,200 --> 00:36:42,480
wasteful to have a really big

835
00:36:39,599 --> 00:36:45,640
Transformer that could solve you know

836
00:36:42,480 --> 00:36:49,119
really complex math problems in the same

837
00:36:45,640 --> 00:36:53,359
amount of time it takes to predict that

838
00:36:49,119 --> 00:36:55,280
the next word is like uh dog after the

839
00:36:53,359 --> 00:36:57,280
word the big or something like that

840
00:36:55,280 --> 00:36:58,560
right so there are some things that are

841
00:36:57,280 --> 00:37:00,000
easy we can do in a second there are

842
00:36:58,560 --> 00:37:01,839
some things that take us more time and

843
00:37:00,000 --> 00:37:05,880
essentially this Chain of Thought

844
00:37:01,839 --> 00:37:09,280
reasoning is um is doing that it's

845
00:37:05,880 --> 00:37:12,280
giving it more time to solve the harder

846
00:37:09,280 --> 00:37:12,280
problems

847
00:37:17,200 --> 00:37:22,440
yes

848
00:37:18,839 --> 00:37:23,960
okay yeah good good question so so

849
00:37:22,440 --> 00:37:26,200
that's what um that's what this next

850
00:37:23,960 --> 00:37:27,920
paper does so uh the the question was

851
00:37:26,200 --> 00:37:31,160
what what happens if we just ask it to

852
00:37:27,920 --> 00:37:34,800
reason and the answer is it still works

853
00:37:31,160 --> 00:37:37,000
um and this paper was really like I I I

854
00:37:34,800 --> 00:37:39,760
love this paper for its Simplicity and

855
00:37:37,000 --> 00:37:43,160
cleverness and basically uh they

856
00:37:39,760 --> 00:37:45,000
contrast few shot learning few shot

857
00:37:43,160 --> 00:37:49,200
Chain of Thought where you provide Chain

858
00:37:45,000 --> 00:37:52,160
of Thought examples zero shot prompting

859
00:37:49,200 --> 00:37:54,560
basically and zero shot Chain of Thought

860
00:37:52,160 --> 00:37:58,720
So what they do is they just

861
00:37:54,560 --> 00:38:00,280
add uh let's thinks step by step that

862
00:37:58,720 --> 00:38:04,200
they add that phrase to the end of The

863
00:38:00,280 --> 00:38:06,079
Prompt and then that elicits the model

864
00:38:04,200 --> 00:38:08,000
to basically do Chain of Thought

865
00:38:06,079 --> 00:38:09,240
reasoning without any further examples

866
00:38:08,000 --> 00:38:12,599
of how that Chain of Thought reasoning

867
00:38:09,240 --> 00:38:14,440
works why does this work again because

868
00:38:12,599 --> 00:38:16,760
like on the internet there's a bunch of

869
00:38:14,440 --> 00:38:20,240
examples of math problem solving data

870
00:38:16,760 --> 00:38:22,800
sets or QA corpora where it says let

871
00:38:20,240 --> 00:38:24,480
things step by step and after that you

872
00:38:22,800 --> 00:38:28,040
you know consistently have this sort of

873
00:38:24,480 --> 00:38:29,800
resoning chain added there so um good

874
00:38:28,040 --> 00:38:31,200
good intuition uh that this paper

875
00:38:29,800 --> 00:38:32,480
answers the question and this like

876
00:38:31,200 --> 00:38:36,119
actually does

877
00:38:32,480 --> 00:38:39,119
work one interesting thing is

878
00:38:36,119 --> 00:38:39,119
um

879
00:38:39,720 --> 00:38:45,200
now if I go to chat

880
00:38:47,319 --> 00:38:52,240
GPT and I say

881
00:38:50,480 --> 00:38:58,520
um

882
00:38:52,240 --> 00:38:58,520
I am teaching a class with 98

883
00:38:58,720 --> 00:39:06,000
students

884
00:39:01,480 --> 00:39:08,400
70% turn in the

885
00:39:06,000 --> 00:39:10,720
assignment hint on

886
00:39:08,400 --> 00:39:17,720
time uh

887
00:39:10,720 --> 00:39:21,880
10% and it in play how many did not ENT

888
00:39:17,720 --> 00:39:21,880
in let's see let's see if this

889
00:39:25,079 --> 00:39:29,200
works okay it's writing code for

890
00:39:29,440 --> 00:39:34,599
me which is that's a feature slide I

891
00:39:32,119 --> 00:39:37,319
kind of didn't uh I kind of didn't

892
00:39:34,599 --> 00:39:37,319
wanted to do

893
00:39:40,040 --> 00:39:44,160
that okay

894
00:39:45,920 --> 00:39:50,359
um I do not

895
00:39:54,280 --> 00:39:59,720
like okay so um

896
00:39:57,040 --> 00:39:59,720
it's a little bit

897
00:40:15,720 --> 00:40:21,839
slow okay so there there that worked but

898
00:40:19,680 --> 00:40:24,640
not that I did not say let's think step

899
00:40:21,839 --> 00:40:27,760
by step I didn't I didn't ask it to do

900
00:40:24,640 --> 00:40:29,240
this um and the reason why is um we're

901
00:40:27,760 --> 00:40:31,119
going to talk about instruction tuning

902
00:40:29,240 --> 00:40:34,359
next time but basically GPT has been

903
00:40:31,119 --> 00:40:36,560
tuned to do this reasoning even if you

904
00:40:34,359 --> 00:40:38,480
don't ask it to do that uh it wouldn't

905
00:40:36,560 --> 00:40:40,839
do that naturally but it's because lots

906
00:40:38,480 --> 00:40:43,880
of supervised data has been added into

907
00:40:40,839 --> 00:40:46,920
this model so like another thing is like

908
00:40:43,880 --> 00:40:48,960
if you are planning on doing anything

909
00:40:46,920 --> 00:40:51,240
about like Chain of Thought reasoning or

910
00:40:48,960 --> 00:40:53,000
or stuff like that as a class project

911
00:40:51,240 --> 00:40:54,960
you need to keep in mind that the like

912
00:40:53,000 --> 00:40:58,280
GPD models have already been trained to

913
00:40:54,960 --> 00:41:00,040
do this and so so if you want to like

914
00:40:58,280 --> 00:41:01,599
try to find out a better way to elicit

915
00:41:00,040 --> 00:41:03,960
this from a raw model you'll need to use

916
00:41:01,599 --> 00:41:07,119
a raw model like llama 2 with no chat

917
00:41:03,960 --> 00:41:10,200
tuning or stuff like that um in order to

918
00:41:07,119 --> 00:41:12,520
uh do that in a neutral in a neutral

919
00:41:10,200 --> 00:41:14,960
setting that hasn't been contaminated by

920
00:41:12,520 --> 00:41:14,960
like super

921
00:41:15,960 --> 00:41:20,520
L cool um any

922
00:41:21,079 --> 00:41:27,720
questions okay um so next I want to talk

923
00:41:24,720 --> 00:41:31,280
about prompting in programs that uh Chad

924
00:41:27,720 --> 00:41:35,560
GPD gave me a good example of uh why why

925
00:41:31,280 --> 00:41:37,160
this is useful or important um so

926
00:41:35,560 --> 00:41:40,640
there's two results actually both of

927
00:41:37,160 --> 00:41:43,440
these are are from my uh collaborators

928
00:41:40,640 --> 00:41:45,839
but the first one is um it demonstrates

929
00:41:43,440 --> 00:41:48,720
that structuring outputs at programs can

930
00:41:45,839 --> 00:41:51,599
help you get better results even if the

931
00:41:48,720 --> 00:41:55,119
task isn't a programmatic task so this

932
00:41:51,599 --> 00:41:57,000
is kind of interesting um so we were

933
00:41:55,119 --> 00:41:59,319
looking at predicting stru structured

934
00:41:57,000 --> 00:42:01,640
outputs and these structured outputs

935
00:41:59,319 --> 00:42:03,839
specifically are procedural knowledge

936
00:42:01,640 --> 00:42:06,920
like this so like how do we cook a pie

937
00:42:03,839 --> 00:42:09,040
or how do we serve pot pies on a plate

938
00:42:06,920 --> 00:42:10,800
and we had this procedural knowledge

939
00:42:09,040 --> 00:42:14,040
like take the pies out to pool open the

940
00:42:10,800 --> 00:42:16,079
cabinet drawer take out several plates

941
00:42:14,040 --> 00:42:17,720
and we wanted to know the dependencies

942
00:42:16,079 --> 00:42:19,520
between these so we could create a

943
00:42:17,720 --> 00:42:22,559
structured like procedural knowledge

944
00:42:19,520 --> 00:42:25,599
base so this is not an inherently code

945
00:42:22,559 --> 00:42:27,200
based task it's not a you know you could

946
00:42:25,599 --> 00:42:28,880
just ask

947
00:42:27,200 --> 00:42:32,160
the model in natural language and that

948
00:42:28,880 --> 00:42:35,000
would work as well so we structured

949
00:42:32,160 --> 00:42:37,720
things in a couple varieties so we had a

950
00:42:35,000 --> 00:42:39,800
textual format we had uh something in

951
00:42:37,720 --> 00:42:43,079
the dot format which is a way to draw

952
00:42:39,800 --> 00:42:45,480
graphs and then we had we also tried

953
00:42:43,079 --> 00:42:47,240
structuring the output in Python so

954
00:42:45,480 --> 00:42:48,960
these are just different ways to format

955
00:42:47,240 --> 00:42:50,720
the output they all say the same thing

956
00:42:48,960 --> 00:42:54,599
and we can extract the answer from all

957
00:42:50,720 --> 00:42:56,920
of them um but we found that structuring

958
00:42:54,599 --> 00:42:58,480
it in in Python basically is the more

959
00:42:56,920 --> 00:43:02,920
effective way of doing

960
00:42:58,480 --> 00:43:04,680
this so why why is it this the case the

961
00:43:02,920 --> 00:43:06,280
answer is essentially the same thing

962
00:43:04,680 --> 00:43:08,680
that I was talking about before with you

963
00:43:06,280 --> 00:43:11,480
know predicting excellent instead of

964
00:43:08,680 --> 00:43:13,319
five right you know it's seen a ton of

965
00:43:11,480 --> 00:43:15,960
python in it's training data so it's

966
00:43:13,319 --> 00:43:17,760
very good at predicting python uh it's

967
00:43:15,960 --> 00:43:20,359
less good at predicting dot format

968
00:43:17,760 --> 00:43:24,240
because it seemed less do format and it

969
00:43:20,359 --> 00:43:26,640
hasn't seen very much text here

970
00:43:24,240 --> 00:43:29,960
um another

971
00:43:26,640 --> 00:43:32,359
comment is code is very highly

972
00:43:29,960 --> 00:43:33,559
structured compared to natural language

973
00:43:32,359 --> 00:43:35,599
and because code is very highly

974
00:43:33,559 --> 00:43:37,520
structured we have things like

975
00:43:35,599 --> 00:43:39,079
dependencies where we refer back to

976
00:43:37,520 --> 00:43:41,079
variables that we defined before and

977
00:43:39,079 --> 00:43:44,119
other things like this so I think when

978
00:43:41,079 --> 00:43:46,760
it starts outputting code the models get

979
00:43:44,119 --> 00:43:48,359
into this mode which say yes please

980
00:43:46,760 --> 00:43:51,280
refer back to the things you've seen

981
00:43:48,359 --> 00:43:53,440
previously more often like attend to

982
00:43:51,280 --> 00:43:57,040
previous stuff more often and don't just

983
00:43:53,440 --> 00:43:59,760
like generate things you know uh

984
00:43:57,040 --> 00:44:02,440
arbitrarily and hallucinate you know new

985
00:43:59,760 --> 00:44:04,119
content and because of this for

986
00:44:02,440 --> 00:44:05,559
generating structured outputs even if

987
00:44:04,119 --> 00:44:08,920
the structured outputs don't need to be

988
00:44:05,559 --> 00:44:11,520
code you can benefit by doing

989
00:44:08,920 --> 00:44:13,200
this another thing that's a really handy

990
00:44:11,520 --> 00:44:16,319
trick is anytime you want to get a

991
00:44:13,200 --> 00:44:19,079
structured output out of a model

992
00:44:16,319 --> 00:44:22,760
um you can ask it to generate something

993
00:44:19,079 --> 00:44:24,839
in Json instead of generating it in uh

994
00:44:22,760 --> 00:44:26,640
in text and the reason why Json is

995
00:44:24,839 --> 00:44:28,079
useful is you can press the on you can

996
00:44:26,640 --> 00:44:32,319
pull out the strings and other stuff

997
00:44:28,079 --> 00:44:34,839
like that um so this can be very

998
00:44:32,319 --> 00:44:37,960
effective because if you just add an

999
00:44:34,839 --> 00:44:40,200
instruction that says please um please

1000
00:44:37,960 --> 00:44:42,440
format things in this particular

1001
00:44:40,200 --> 00:44:43,680
format often the model won't listen to

1002
00:44:42,440 --> 00:44:44,800
you and it will output something in a

1003
00:44:43,680 --> 00:44:46,280
different format you need to write a

1004
00:44:44,800 --> 00:44:48,599
really annoying parser to pull out the

1005
00:44:46,280 --> 00:44:50,280
information that you actually want but

1006
00:44:48,599 --> 00:44:51,960
it gets Json right almost all of the

1007
00:44:50,280 --> 00:44:54,040
time just because it's seen so much Json

1008
00:44:51,960 --> 00:44:57,880
so that's a nice trick if you want to do

1009
00:44:54,040 --> 00:45:01,520
something like that

1010
00:44:57,880 --> 00:45:03,559
another uh thing is a paper uh called

1011
00:45:01,520 --> 00:45:08,079
program AED language models that we did

1012
00:45:03,559 --> 00:45:10,200
about a year ago and the method that we

1013
00:45:08,079 --> 00:45:13,760
proposed here is using a program to

1014
00:45:10,200 --> 00:45:16,480
generate outputs uh using a program to

1015
00:45:13,760 --> 00:45:19,440
generate outputs and this can be more

1016
00:45:16,480 --> 00:45:22,319
precise than asking an LM to do so and

1017
00:45:19,440 --> 00:45:26,720
so instead of doing Chain of Thought

1018
00:45:22,319 --> 00:45:30,640
prompting we created a few F shot

1019
00:45:26,720 --> 00:45:34,319
examples where we wrote like the text

1020
00:45:30,640 --> 00:45:37,160
here and then the text in English and

1021
00:45:34,319 --> 00:45:40,160
then we had code corresponding code the

1022
00:45:37,160 --> 00:45:42,280
text in English corresponding code and

1023
00:45:40,160 --> 00:45:44,960
then the answer is and then the final

1024
00:45:42,280 --> 00:45:48,160
code and then we basically generate this

1025
00:45:44,960 --> 00:45:49,640
code and execute it to get the answer so

1026
00:45:48,160 --> 00:45:52,319
like as you saw this is implemented in

1027
00:45:49,640 --> 00:45:54,280
chat GP now it's uh you write something

1028
00:45:52,319 --> 00:45:56,319
out it will decide whether it wants to

1029
00:45:54,280 --> 00:45:58,599
generate code or generate text depending

1030
00:45:56,319 --> 00:46:00,760
on the type of problem and it's just

1031
00:45:58,599 --> 00:46:03,559
more precise it can solve like actually

1032
00:46:00,760 --> 00:46:05,200
rather complex problems like uh you know

1033
00:46:03,559 --> 00:46:07,880
calculating how much tax you need to be

1034
00:46:05,200 --> 00:46:10,880
paying or something like that

1035
00:46:07,880 --> 00:46:12,480
um it's especially useful for numeric

1036
00:46:10,880 --> 00:46:14,599
questions and it's implemented in things

1037
00:46:12,480 --> 00:46:17,040
like the chat GPT uh code interpreter

1038
00:46:14,599 --> 00:46:18,640
bar to execution other things like that

1039
00:46:17,040 --> 00:46:22,079
it's pretty cool it can actually do

1040
00:46:18,640 --> 00:46:24,440
visualizations for you for papers also

1041
00:46:22,079 --> 00:46:28,000
so if you ask it to visualize data for

1042
00:46:24,440 --> 00:46:30,200
you um chat GPD now does a pretty good

1043
00:46:28,000 --> 00:46:32,640
job of doing this like to give an

1044
00:46:30,200 --> 00:46:34,319
example I asked it I gave it a big

1045
00:46:32,640 --> 00:46:35,760
python list and asked it to generate a

1046
00:46:34,319 --> 00:46:37,839
histogram and it did a really good job

1047
00:46:35,760 --> 00:46:40,240
of it for me it also gives you the code

1048
00:46:37,839 --> 00:46:42,839
so you can go in and modify it later so

1049
00:46:40,240 --> 00:46:44,720
um I would definitely recommend you know

1050
00:46:42,839 --> 00:46:46,200
thinking about using this uh either in

1051
00:46:44,720 --> 00:46:49,200
your research or just to write your

1052
00:46:46,200 --> 00:46:51,480
reports uh for this class so um this

1053
00:46:49,200 --> 00:46:55,839
class is uh generative AI friendly

1054
00:46:51,480 --> 00:46:57,760
mostly so like I do I do want you to

1055
00:46:55,839 --> 00:46:59,880
learn the things we expect you to learn

1056
00:46:57,760 --> 00:47:02,480
which is why I suggest that you don't

1057
00:46:59,880 --> 00:47:04,400
like just write every uh everything for

1058
00:47:02,480 --> 00:47:06,280
assignment number one with chat GP key

1059
00:47:04,400 --> 00:47:07,720
but I think even if you tried to do that

1060
00:47:06,280 --> 00:47:09,640
it'd probably get it wrong in subtle

1061
00:47:07,720 --> 00:47:10,920
ways so you're probably better off

1062
00:47:09,640 --> 00:47:13,880
understanding the content

1063
00:47:10,920 --> 00:47:16,400
anyway um

1064
00:47:13,880 --> 00:47:18,160
cool this can also be expanded a whole

1065
00:47:16,400 --> 00:47:21,559
lot into like agents and tools and I'm

1066
00:47:18,160 --> 00:47:21,559
going to talk about that separately

1067
00:47:22,800 --> 00:47:27,720
later cool uh any any things about this

1068
00:47:29,040 --> 00:47:34,200
okay I'm uh I'm going to go next so

1069
00:47:31,800 --> 00:47:36,079
prompt engineering um when you're

1070
00:47:34,200 --> 00:47:37,280
designing prompts uh there's a number of

1071
00:47:36,079 --> 00:47:38,240
different ways you can do this you can

1072
00:47:37,280 --> 00:47:41,559
do this

1073
00:47:38,240 --> 00:47:42,960
manually uh you to do this you configure

1074
00:47:41,559 --> 00:47:44,520
a manual template based on the

1075
00:47:42,960 --> 00:47:46,160
characteristics of the task using all of

1076
00:47:44,520 --> 00:47:48,880
the knowledge that I told you

1077
00:47:46,160 --> 00:47:50,119
before you can also do automated search

1078
00:47:48,880 --> 00:47:52,079
and there's a number of different ways

1079
00:47:50,119 --> 00:47:55,119
to do automated search for

1080
00:47:52,079 --> 00:47:58,319
prompts uh the first one is doing some

1081
00:47:55,119 --> 00:48:00,599
sort of search discret space uh so you

1082
00:47:58,319 --> 00:48:02,720
find a prompt that is

1083
00:48:00,599 --> 00:48:04,680
essentially

1084
00:48:02,720 --> 00:48:06,640
text the other one is search in

1085
00:48:04,680 --> 00:48:08,559
continuous space so you find a prompt

1086
00:48:06,640 --> 00:48:10,680
that isn't actually comprehensible text

1087
00:48:08,559 --> 00:48:14,760
but nonetheless is a good

1088
00:48:10,680 --> 00:48:16,960
prompt so looking at manual engineering

1089
00:48:14,760 --> 00:48:19,000
um making sure that the format matches

1090
00:48:16,960 --> 00:48:21,680
that of a trained model uh such as the

1091
00:48:19,000 --> 00:48:24,359
chat format is actually really really

1092
00:48:21,680 --> 00:48:26,119
important um and this can have a a large

1093
00:48:24,359 --> 00:48:28,119
effect on models there's a really paper

1094
00:48:26,119 --> 00:48:30,000
that demonstrated this convincingly

1095
00:48:28,119 --> 00:48:33,200
before and also releases some software

1096
00:48:30,000 --> 00:48:35,880
that allows you to do this um kind of in

1097
00:48:33,200 --> 00:48:38,079
an efficient manner and what this is

1098
00:48:35,880 --> 00:48:41,200
showing is

1099
00:48:38,079 --> 00:48:45,079
um this is the original formatting of a

1100
00:48:41,200 --> 00:48:48,400
prompt that was given I I Believe by uh

1101
00:48:45,079 --> 00:48:50,119
some sort of like uh machine reading or

1102
00:48:48,400 --> 00:48:52,799
document based question answering data

1103
00:48:50,119 --> 00:48:55,480
set which was like passage

1104
00:48:52,799 --> 00:48:58,440
answer if you modify the spacing between

1105
00:48:55,480 --> 00:49:01,680
the the fields that increases your score

1106
00:48:58,440 --> 00:49:04,280
by several percentage points um if you

1107
00:49:01,680 --> 00:49:06,880
remove the colons that increases your

1108
00:49:04,280 --> 00:49:08,720
score by a few more percentage points

1109
00:49:06,880 --> 00:49:10,119
it's kind of silly but like little

1110
00:49:08,720 --> 00:49:11,040
things like this actually can make a

1111
00:49:10,119 --> 00:49:14,240
really big

1112
00:49:11,040 --> 00:49:17,599
difference um if you modify the casing

1113
00:49:14,240 --> 00:49:19,960
this decreases by a lot if you modify

1114
00:49:17,599 --> 00:49:22,440
the casing and remove colons so the

1115
00:49:19,960 --> 00:49:25,200
thing that was useful like adding colons

1116
00:49:22,440 --> 00:49:26,720
here remove colons uh that further

1117
00:49:25,200 --> 00:49:29,280
decrease

1118
00:49:26,720 --> 00:49:31,400
if you forget to add a space between the

1119
00:49:29,280 --> 00:49:32,559
passage and the text that really hurts

1120
00:49:31,400 --> 00:49:35,599
your

1121
00:49:32,559 --> 00:49:38,000
accuracy so this is pretty painful right

1122
00:49:35,599 --> 00:49:40,599
like you don't want to be getting uh

1123
00:49:38,000 --> 00:49:44,160
0.036% accuracy when adding a space

1124
00:49:40,599 --> 00:49:48,680
would give you like 75% accuracy

1125
00:49:44,160 --> 00:49:50,799
right um and one interesting thing is um

1126
00:49:48,680 --> 00:49:53,160
this is looking

1127
00:49:50,799 --> 00:49:56,559
at different

1128
00:49:53,160 --> 00:49:58,520
models and um

1129
00:49:56,559 --> 00:50:00,640
with different models it's pretty

1130
00:49:58,520 --> 00:50:03,599
consistent that many different plausible

1131
00:50:00,640 --> 00:50:05,400
formats that you try the average gives

1132
00:50:03,599 --> 00:50:07,240
you a really low accuracy but there's a

1133
00:50:05,400 --> 00:50:08,760
few outliers that give you really good

1134
00:50:07,240 --> 00:50:11,119
accuracy and these probably correspond

1135
00:50:08,760 --> 00:50:13,400
to the things that it was trained on um

1136
00:50:11,119 --> 00:50:15,880
instruction tuned on or or other things

1137
00:50:13,400 --> 00:50:17,480
like this so number one make sure you're

1138
00:50:15,880 --> 00:50:19,799
using like the canonical prompt

1139
00:50:17,480 --> 00:50:21,240
formatting for the model for sure number

1140
00:50:19,799 --> 00:50:22,640
two you might want to do a little bit of

1141
00:50:21,240 --> 00:50:24,720
additional search to see if you can do

1142
00:50:22,640 --> 00:50:26,960
even better than that so um this is

1143
00:50:24,720 --> 00:50:29,480
something to be very aware

1144
00:50:26,960 --> 00:50:32,480
of

1145
00:50:29,480 --> 00:50:32,480
um

1146
00:50:34,200 --> 00:50:37,680
okay do you have a

1147
00:50:39,599 --> 00:50:43,720
question this is dependent on what it

1148
00:50:41,720 --> 00:50:47,680
sees in trading time another thing

1149
00:50:43,720 --> 00:50:51,920
actually is um this will definitely be

1150
00:50:47,680 --> 00:50:53,200
tighter for uh like a chat GPT or GPT 4

1151
00:50:51,920 --> 00:50:56,599
um because it's been trained on many

1152
00:50:53,200 --> 00:50:59,319
different formats at training time um

1153
00:50:56,599 --> 00:51:00,880
and so the better the model has been

1154
00:50:59,319 --> 00:51:03,520
trained on a lot of different formats

1155
00:51:00,880 --> 00:51:05,559
the less this is going to have an

1156
00:51:03,520 --> 00:51:06,920
effect but you know you're probably not

1157
00:51:05,559 --> 00:51:09,440
going to be retraining a model that

1158
00:51:06,920 --> 00:51:10,799
somebody gives you uh so like this is

1159
00:51:09,440 --> 00:51:12,880
something to be very aware of if you're

1160
00:51:10,799 --> 00:51:14,839
just a downstream newer model especially

1161
00:51:12,880 --> 00:51:17,599
an open source

1162
00:51:14,839 --> 00:51:19,359
model um another thing is how do you

1163
00:51:17,599 --> 00:51:22,280
give instructions to

1164
00:51:19,359 --> 00:51:25,000
models um instructions should be clear

1165
00:51:22,280 --> 00:51:29,280
concise and easy to understand one very

1166
00:51:25,000 --> 00:51:31,559
funny thing is um I think now like

1167
00:51:29,280 --> 00:51:33,280
actually prompting language models is

1168
00:51:31,559 --> 00:51:34,960
very similar to prompting humans

1169
00:51:33,280 --> 00:51:37,119
especially if we're talking about like

1170
00:51:34,960 --> 00:51:38,760
gp4 so if you're not very good at

1171
00:51:37,119 --> 00:51:41,599
explaining things to humans that might

1172
00:51:38,760 --> 00:51:45,440
actually be bad um and you might want to

1173
00:51:41,599 --> 00:51:47,359
practice that and explaining things to

1174
00:51:45,440 --> 00:51:50,319
models might be a good way to practice

1175
00:51:47,359 --> 00:51:51,799
that right so you know um it actually

1176
00:51:50,319 --> 00:51:54,040
can give you feedback without annoying

1177
00:51:51,799 --> 00:51:55,359
your friends by having you explain uh

1178
00:51:54,040 --> 00:51:58,160
things to them in several different ways

1179
00:51:55,359 --> 00:52:00,040
way and seeing how they react so um but

1180
00:51:58,160 --> 00:52:03,680
anyway clear concise easy to understand

1181
00:52:00,040 --> 00:52:05,319
is good um there's this prompting guide

1182
00:52:03,680 --> 00:52:08,599
uh which I I can

1183
00:52:05,319 --> 00:52:13,240
open um this has a prompt engineering

1184
00:52:08,599 --> 00:52:14,520
guide I I I like this site but it it

1185
00:52:13,240 --> 00:52:17,400
does have a bit

1186
00:52:14,520 --> 00:52:18,880
of like variance in the importance of

1187
00:52:17,400 --> 00:52:21,760
the information that tells you but like

1188
00:52:18,880 --> 00:52:23,960
this particular page is nice I feel so

1189
00:52:21,760 --> 00:52:26,160
start simple start with simple

1190
00:52:23,960 --> 00:52:29,520
instructions um

1191
00:52:26,160 --> 00:52:32,119
you should tell the model what it should

1192
00:52:29,520 --> 00:52:36,839
be doing so make sure you say write

1193
00:52:32,119 --> 00:52:39,799
classify summarize translate order um

1194
00:52:36,839 --> 00:52:41,960
and things like this uh it also gives

1195
00:52:39,799 --> 00:52:45,440
some good examples of the level of

1196
00:52:41,960 --> 00:52:47,559
specificity that you should be giving so

1197
00:52:45,440 --> 00:52:49,680
something that's less precise is explain

1198
00:52:47,559 --> 00:52:51,559
the concept of prompt engineering keep

1199
00:52:49,680 --> 00:52:53,920
the explanation short only a few

1200
00:52:51,559 --> 00:52:57,119
sentences and don't be too

1201
00:52:53,920 --> 00:52:58,799
descriptive um it use two to three

1202
00:52:57,119 --> 00:53:00,240
sentences to explain the concept of

1203
00:52:58,799 --> 00:53:02,599
prompt engineering to a high school

1204
00:53:00,240 --> 00:53:04,839
student so what this does is this tells

1205
00:53:02,599 --> 00:53:07,839
you the level of read like the reading

1206
00:53:04,839 --> 00:53:07,839
level

1207
00:53:07,960 --> 00:53:12,520
um so this doesn't even tell you the

1208
00:53:10,200 --> 00:53:14,319
reading level I guess um and then two to

1209
00:53:12,520 --> 00:53:16,240
three sentences is more precise than

1210
00:53:14,319 --> 00:53:19,200
keep it a few sentences don't be too

1211
00:53:16,240 --> 00:53:22,440
descriptive so um the more precise you

1212
00:53:19,200 --> 00:53:25,760
can be the the better it is um one

1213
00:53:22,440 --> 00:53:27,040
interesting thing is like if you ask

1214
00:53:25,760 --> 00:53:28,359
your friend to do something and they

1215
00:53:27,040 --> 00:53:32,400
don't know how to do it they'll complain

1216
00:53:28,359 --> 00:53:34,240
to you but right now uh LMS don't

1217
00:53:32,400 --> 00:53:35,720
complain to you they may in the future

1218
00:53:34,240 --> 00:53:38,680
uh that might be like actually an

1219
00:53:35,720 --> 00:53:40,799
interesting thing to find uh the you

1220
00:53:38,680 --> 00:53:42,319
know interesting methodological thing to

1221
00:53:40,799 --> 00:53:45,240
look at for a project or something like

1222
00:53:42,319 --> 00:53:47,960
that but um right now you need to be

1223
00:53:45,240 --> 00:53:49,040
precise and like there's it doesn't give

1224
00:53:47,960 --> 00:53:51,799
you feedback when you're not being

1225
00:53:49,040 --> 00:53:51,799
precise so you need

1226
00:53:52,000 --> 00:53:56,359
to um separately from this there are

1227
00:53:54,200 --> 00:53:59,160
methods for automatic prompt engineering

1228
00:53:56,359 --> 00:54:00,960
so uh prompt paraphrasing gradient based

1229
00:53:59,160 --> 00:54:02,240
discreet prompt search prompt tuning

1230
00:54:00,960 --> 00:54:06,160
prefix

1231
00:54:02,240 --> 00:54:09,880
tuning so prompt paraphrasing um this is

1232
00:54:06,160 --> 00:54:12,559
a method that uh we proposed a while ago

1233
00:54:09,880 --> 00:54:15,760
um to basically paraphrase an existing

1234
00:54:12,559 --> 00:54:17,280
prompt to get other candidates um it's

1235
00:54:15,760 --> 00:54:19,240
rather simple basically you take a

1236
00:54:17,280 --> 00:54:21,960
prompt you put it through a paraphrasing

1237
00:54:19,240 --> 00:54:24,280
model and it will give you new prompts

1238
00:54:21,960 --> 00:54:25,440
and this is good because it will tend to

1239
00:54:24,280 --> 00:54:28,319
give you things that are natural

1240
00:54:25,440 --> 00:54:29,839
language um you can paraphrase 50 times

1241
00:54:28,319 --> 00:54:32,480
try all of them see which one gives you

1242
00:54:29,839 --> 00:54:37,079
the highest accuracy and then use that

1243
00:54:32,480 --> 00:54:39,280
one um there's also an interesting paper

1244
00:54:37,079 --> 00:54:43,079
uh that demonstrates that you can do

1245
00:54:39,280 --> 00:54:45,240
this iteratively so you paraphrase once

1246
00:54:43,079 --> 00:54:46,599
um you filter down all the candidates

1247
00:54:45,240 --> 00:54:48,119
that do well and then you go in and

1248
00:54:46,599 --> 00:54:49,960
paraphrase them again and you just do

1249
00:54:48,119 --> 00:54:51,960
this over and over again and that can

1250
00:54:49,960 --> 00:54:54,079
give you better results than kind of one

1251
00:54:51,960 --> 00:54:57,079
one off

1252
00:54:54,079 --> 00:54:57,079
paraphrasing

1253
00:54:59,240 --> 00:55:02,079
so that's very simple you can even use a

1254
00:55:01,079 --> 00:55:04,160
large language model to do the

1255
00:55:02,079 --> 00:55:06,599
paraphrasing for you um another thing

1256
00:55:04,160 --> 00:55:08,920
that you can do is gradient based search

1257
00:55:06,599 --> 00:55:11,119
so the way this works is you need to

1258
00:55:08,920 --> 00:55:16,319
have a a model that you can calculate

1259
00:55:11,119 --> 00:55:19,920
gradients for and what you do is you

1260
00:55:16,319 --> 00:55:22,240
calculate you create a seed prompt and

1261
00:55:19,920 --> 00:55:26,000
then you calculate gradients into that

1262
00:55:22,240 --> 00:55:29,760
seed prompt so you treat the

1263
00:55:26,000 --> 00:55:33,160
um you treat each of the tokens here

1264
00:55:29,760 --> 00:55:36,680
like T1 T2 T3 T4

1265
00:55:33,160 --> 00:55:38,240
T5 as their own embeddings you do back

1266
00:55:36,680 --> 00:55:39,920
propop into those embeddings and you

1267
00:55:38,240 --> 00:55:42,799
optimize them to get high accuracy on

1268
00:55:39,920 --> 00:55:44,720
your data set then after you're done

1269
00:55:42,799 --> 00:55:47,319
optimizing them to get high accuracy on

1270
00:55:44,720 --> 00:55:49,079
your data set you clamp them onto the

1271
00:55:47,319 --> 00:55:52,160
nearest neighbor embedding that you

1272
00:55:49,079 --> 00:55:53,520
already have so you basically say okay

1273
00:55:52,160 --> 00:55:56,720
the nearest neighbor to the embedding

1274
00:55:53,520 --> 00:55:58,920
that I learned you um is atmosphere then

1275
00:55:56,720 --> 00:56:02,240
a lot dialogue clone

1276
00:55:58,920 --> 00:56:03,799
totally and so this is this will

1277
00:56:02,240 --> 00:56:05,599
actually give you better results than

1278
00:56:03,799 --> 00:56:07,839
paraphrasing in many cases because the

1279
00:56:05,599 --> 00:56:11,520
search space is less constrained you can

1280
00:56:07,839 --> 00:56:12,960
get these very unnatural prompts uh that

1281
00:56:11,520 --> 00:56:16,280
don't seem to make sense but actually

1282
00:56:12,960 --> 00:56:20,280
work well this has particularly been

1283
00:56:16,280 --> 00:56:22,960
widely used in um adversarial attacks on

1284
00:56:20,280 --> 00:56:25,599
language modals so how can you come up

1285
00:56:22,960 --> 00:56:27,720
with um

1286
00:56:25,599 --> 00:56:31,559
with prompts that

1287
00:56:27,720 --> 00:56:33,319
cause language models to uh do things

1288
00:56:31,559 --> 00:56:36,039
that you don't want them to be

1289
00:56:33,319 --> 00:56:38,920
doing and um there's actually this nice

1290
00:56:36,039 --> 00:56:41,440
paper uh also by people at CMU called

1291
00:56:38,920 --> 00:56:42,960
Universal and transferable adversarial

1292
00:56:41,440 --> 00:56:45,400
attacks on line language

1293
00:56:42,960 --> 00:56:50,559
models and basically what they do is

1294
00:56:45,400 --> 00:56:53,880
they try to optimize the uh they try to

1295
00:56:50,559 --> 00:56:56,839
optimize the prompt to create a prompt

1296
00:56:53,880 --> 00:56:58,599
that causes the model to do bad things

1297
00:56:56,839 --> 00:57:00,039
basically and they try to do it even on

1298
00:56:58,599 --> 00:57:03,440
models that have been trying to not do

1299
00:57:00,039 --> 00:57:05,039
bad things and they demonstrate that

1300
00:57:03,440 --> 00:57:07,359
number one you can cause things like

1301
00:57:05,039 --> 00:57:09,599
models like llama to do bad you know bad

1302
00:57:07,359 --> 00:57:12,559
things like output toxic things tell you

1303
00:57:09,599 --> 00:57:15,599
how to build bombs stuff like that but

1304
00:57:12,559 --> 00:57:18,480
also the same prompts also work on like

1305
00:57:15,599 --> 00:57:22,319
GPD models uh which is kind of like

1306
00:57:18,480 --> 00:57:23,839
interesting and and very uh you know

1307
00:57:22,319 --> 00:57:26,520
confusing in a way because you thought

1308
00:57:23,839 --> 00:57:28,160
this might be explo idiosyncrasies of a

1309
00:57:26,520 --> 00:57:32,440
particular language model but actually

1310
00:57:28,160 --> 00:57:32,440
it's not so I I find this kind of

1311
00:57:33,880 --> 00:57:39,520
fascinating

1312
00:57:36,039 --> 00:57:42,240
so if you take that a step further one

1313
00:57:39,520 --> 00:57:44,079
thing that you can do is you can say oh

1314
00:57:42,240 --> 00:57:46,280
actually there's no reason why we need

1315
00:57:44,079 --> 00:57:48,520
to clamp these embeddings back to an

1316
00:57:46,280 --> 00:57:52,240
existing embedding right so we could

1317
00:57:48,520 --> 00:57:56,079
just optimize the prompts the embeddings

1318
00:57:52,240 --> 00:57:57,720
of the prompts that go for a task and

1319
00:57:56,079 --> 00:58:02,000
not clamp them back to embeddings and

1320
00:57:57,720 --> 00:58:03,599
just keep them as is so um what I mean

1321
00:58:02,000 --> 00:58:07,079
by that is like right here it's

1322
00:58:03,599 --> 00:58:09,160
optimizing T1 T2 T3 T4 T5 and then

1323
00:58:07,079 --> 00:58:11,359
clamping that back to Atmosphere a lot

1324
00:58:09,160 --> 00:58:13,960
dialog clone totally but just keep them

1325
00:58:11,359 --> 00:58:16,160
as is and don't worry about them like

1326
00:58:13,960 --> 00:58:18,039
actually being a token in the model

1327
00:58:16,160 --> 00:58:19,400
because if you have control over your

1328
00:58:18,039 --> 00:58:21,200
model you can just add them as new

1329
00:58:19,400 --> 00:58:25,960
elements in the vocabulary and you're

1330
00:58:21,200 --> 00:58:28,440
fine right so what they demonstrate in

1331
00:58:25,960 --> 00:58:31,520
this paper is that instead of taking

1332
00:58:28,440 --> 00:58:33,440
your 11 billion parameter model and

1333
00:58:31,520 --> 00:58:35,920
training the whole 11 billion parameter

1334
00:58:33,440 --> 00:58:38,359
model for many different tasks on many

1335
00:58:35,920 --> 00:58:40,079
different data sets they just train

1336
00:58:38,359 --> 00:58:42,039
these prompts which are like 20K

1337
00:58:40,079 --> 00:58:44,039
parameters each I I forget how long it

1338
00:58:42,039 --> 00:58:46,280
is it's like 10 tokens or 20 tokens or

1339
00:58:44,039 --> 00:58:48,079
something like that um and train it on

1340
00:58:46,280 --> 00:58:49,640
all of the the data sets here and you

1341
00:58:48,079 --> 00:58:50,680
don't actually need to do multitask

1342
00:58:49,640 --> 00:58:52,200
learning you don't need to train on

1343
00:58:50,680 --> 00:58:53,720
multiple tasks at the same time you can

1344
00:58:52,200 --> 00:58:56,119
just train on a single

1345
00:58:53,720 --> 00:58:58,599
task

1346
00:58:56,119 --> 00:59:01,000
so now let's take that even a step

1347
00:58:58,599 --> 00:59:03,640
further so this is only training the

1348
00:59:01,000 --> 00:59:06,359
embeddings that you input into the model

1349
00:59:03,640 --> 00:59:08,160
there's a method called prefix tuning

1350
00:59:06,359 --> 00:59:10,319
and the way prefix tuning works is

1351
00:59:08,160 --> 00:59:12,280
instead of training only the embeddings

1352
00:59:10,319 --> 00:59:14,799
that go into the model they actually

1353
00:59:12,280 --> 00:59:18,920
train a prefix that you then append to

1354
00:59:14,799 --> 00:59:20,839
every layer of the model so prompt

1355
00:59:18,920 --> 00:59:23,319
tuning basically does this for the first

1356
00:59:20,839 --> 00:59:24,839
layer of the model prefix tuning does

1357
00:59:23,319 --> 00:59:28,400
this for every layer of the model you

1358
00:59:24,839 --> 00:59:30,319
append a prefix uh for every day so it's

1359
00:59:28,400 --> 00:59:32,200
just a more expressive version of

1360
00:59:30,319 --> 00:59:36,119
prompting

1361
00:59:32,200 --> 00:59:40,200
essentially so these are all kinds of

1362
00:59:36,119 --> 00:59:43,680
gradual steps from a human created

1363
00:59:40,200 --> 00:59:47,880
prompt into something that is basically

1364
00:59:43,680 --> 00:59:50,839
training a a prompt or a prefix to the

1365
00:59:47,880 --> 00:59:52,960
model so I I would take questions but

1366
00:59:50,839 --> 00:59:55,200
let me get to the end of this section uh

1367
00:59:52,960 --> 00:59:58,839
also because uh I think there's

1368
00:59:55,200 --> 01:00:00,720
interesting analogies here so in the

1369
00:59:58,839 --> 01:00:02,880
next class I'm going to talk about

1370
01:00:00,720 --> 01:00:04,440
parameter efficient fine-tuning methods

1371
01:00:02,880 --> 01:00:06,960
which is kind of a more

1372
01:00:04,440 --> 01:00:10,000
General it's a more

1373
01:00:06,960 --> 01:00:11,480
General version of prompt tuning or

1374
01:00:10,000 --> 01:00:13,280
prefix tuning there are methods that

1375
01:00:11,480 --> 01:00:15,960
tune a small number of parameters to get

1376
01:00:13,280 --> 01:00:17,400
the model to do something and there's a

1377
01:00:15,960 --> 01:00:18,880
bunch of different parameter efficient

1378
01:00:17,400 --> 01:00:21,520
tuning methods many people may have

1379
01:00:18,880 --> 01:00:23,880
heard of something like Laura uh or

1380
01:00:21,520 --> 01:00:25,440
adapters um I just talked about prefix

1381
01:00:23,880 --> 01:00:28,119
tuning

1382
01:00:25,440 --> 01:00:30,960
so essentially prompt tuning and prefix

1383
01:00:28,119 --> 01:00:33,359
tuning are part of this more General

1384
01:00:30,960 --> 01:00:36,680
class of parameter efficient find tuning

1385
01:00:33,359 --> 01:00:39,240
methods and so what we can say is

1386
01:00:36,680 --> 01:00:41,119
actually prompting is fine-tuning

1387
01:00:39,240 --> 01:00:42,920
prompting is a way of fine-tuning the

1388
01:00:41,119 --> 01:00:46,799
model or getting the model to perform a

1389
01:00:42,920 --> 01:00:49,839
particular task um and we have this

1390
01:00:46,799 --> 01:00:53,720
taxonomy of we have prompts in natural

1391
01:00:49,839 --> 01:00:55,160
language that are created uh by humans

1392
01:00:53,720 --> 01:00:57,240
actually maybe I should say manual

1393
01:00:55,160 --> 01:00:59,559
prompt engineering here this was first

1394
01:00:57,240 --> 01:01:01,480
done in the gpd2 paper where they

1395
01:00:59,559 --> 01:01:04,359
demonstrate that models uh models could

1396
01:01:01,480 --> 01:01:06,200
solve tasks by doing it this way prompt

1397
01:01:04,359 --> 01:01:07,760
paraphrasing is a step up from this

1398
01:01:06,200 --> 01:01:09,799
because it's no longer relying on human

1399
01:01:07,760 --> 01:01:12,680
engineering and you can you know expand

1400
01:01:09,799 --> 01:01:15,280
to a broader set of prompts um it can

1401
01:01:12,680 --> 01:01:17,359
always start with human created prompts

1402
01:01:15,280 --> 01:01:20,240
so it's kind of like broader uh than

1403
01:01:17,359 --> 01:01:21,799
that discrete prompt search doesn't

1404
01:01:20,240 --> 01:01:23,599
necessarily need to rely on a

1405
01:01:21,799 --> 01:01:25,559
paraphrasing model it could rely on like

1406
01:01:23,599 --> 01:01:26,760
gradient-based models or something else

1407
01:01:25,559 --> 01:01:29,240
like that to give you something that's

1408
01:01:26,760 --> 01:01:32,559
not actually natural language uh kind of

1409
01:01:29,240 --> 01:01:35,920
just random tokens continuous prompts or

1410
01:01:32,559 --> 01:01:38,119
prompt tuning is a step above that

1411
01:01:35,920 --> 01:01:41,039
multi-layer continuous prompts or prefix

1412
01:01:38,119 --> 01:01:42,520
tuning is a layer above that parameter

1413
01:01:41,039 --> 01:01:43,520
efficient tuning is more General than

1414
01:01:42,520 --> 01:01:45,359
that and then you have all training

1415
01:01:43,520 --> 01:01:49,160
methods so including fine tuning your

1416
01:01:45,359 --> 01:01:52,680
model and so what are the implications

1417
01:01:49,160 --> 01:01:55,760
of this um I think so a lot of people

1418
01:01:52,680 --> 01:01:58,720
when prompting came out they were like

1419
01:01:55,760 --> 01:02:00,640
prompting methods are very hacky I don't

1420
01:01:58,720 --> 01:02:03,839
like how we have to do manual prompt

1421
01:02:00,640 --> 01:02:08,160
engineering um it seems like a dark art

1422
01:02:03,839 --> 01:02:11,000
as opposed to like you know actually you

1423
01:02:08,160 --> 01:02:14,160
know some sort of well understood

1424
01:02:11,000 --> 01:02:16,839
fine-tuning method that we could use um

1425
01:02:14,160 --> 01:02:20,520
but I I actually like them I like

1426
01:02:16,839 --> 01:02:23,920
prompting a lot because um if anybody is

1427
01:02:20,520 --> 01:02:25,960
familiar with like basian basian

1428
01:02:23,920 --> 01:02:27,920
statistics or machine learning we have

1429
01:02:25,960 --> 01:02:28,799
the concept of like a prior probability

1430
01:02:27,920 --> 01:02:31,200
over

1431
01:02:28,799 --> 01:02:32,359
parameters and then a probability that

1432
01:02:31,200 --> 01:02:34,680
we get

1433
01:02:32,359 --> 01:02:37,880
after after fine tuning the model or

1434
01:02:34,680 --> 01:02:40,440
after training the model and prompts in

1435
01:02:37,880 --> 01:02:42,640
a way are our first like good prior over

1436
01:02:40,440 --> 01:02:43,880
neural network models they give us the

1437
01:02:42,640 --> 01:02:46,319
ability to

1438
01:02:43,880 --> 01:02:48,559
specify what task the model should be

1439
01:02:46,319 --> 01:02:51,880
doing or like a general idea of what

1440
01:02:48,559 --> 01:02:54,200
task the model should be doing before we

1441
01:02:51,880 --> 01:02:56,359
ask the model to actually do the task

1442
01:02:54,200 --> 01:02:58,640
and and so we can either use that prior

1443
01:02:56,359 --> 01:03:02,119
Asis we can use a prompted model Asis

1444
01:02:58,640 --> 01:03:04,839
without doing any additional tuning or

1445
01:03:02,119 --> 01:03:06,480
we could take the prior that we have

1446
01:03:04,839 --> 01:03:07,920
given to the model by using a natural

1447
01:03:06,480 --> 01:03:09,039
language description of the task it

1448
01:03:07,920 --> 01:03:12,079
should be

1449
01:03:09,039 --> 01:03:14,799
doing and then combine it with fineing

1450
01:03:12,079 --> 01:03:17,039
so we can take the prompted

1451
01:03:14,799 --> 01:03:19,279
model we can

1452
01:03:17,039 --> 01:03:21,640
initialize we can initialize the

1453
01:03:19,279 --> 01:03:23,960
distribution of this like Cas a prompt

1454
01:03:21,640 --> 01:03:25,720
using the prompt using a human created

1455
01:03:23,960 --> 01:03:28,160
prompt and then go on and fine-tune it

1456
01:03:25,720 --> 01:03:30,960
on lots of training data as well and

1457
01:03:28,160 --> 01:03:33,799
there's a method for doing that um by

1458
01:03:30,960 --> 01:03:35,880
shik and schutza uh called uh pattern

1459
01:03:33,799 --> 01:03:37,559
exploiting training where they do

1460
01:03:35,880 --> 01:03:39,799
exactly that they basically initialize

1461
01:03:37,559 --> 01:03:41,720
with a manually created prompt and then

1462
01:03:39,799 --> 01:03:44,559
they find the model on finding inator

1463
01:03:41,720 --> 01:03:46,400
after that so um that's a reason why I

1464
01:03:44,559 --> 01:03:47,920
like prompting based methods they they

1465
01:03:46,400 --> 01:03:49,720
give us this like really nice way to

1466
01:03:47,920 --> 01:03:53,039
very quickly create a system but we can

1467
01:03:49,720 --> 01:03:56,079
also have you know whatever level of

1468
01:03:53,039 --> 01:03:59,880
additional training on top of that

1469
01:03:56,079 --> 01:03:59,880
cool so that's a little bit early I'm