Datasets:
jamescalam
/
youtube-transcriptions

Dataset card Data Studio Files
xet
Community
2
title
stringlengths
12
112
published
stringlengths
19
23
url
stringlengths
28
28
video_id
stringlengths
11
11
channel_id
stringclasses
5 values
id
stringlengths
16
31
text
stringlengths
0
596
start
float64
0
37.8k
end
float64
2.18
37.8k
Neural Networks from Scratch - P.3 The Dot Product
2020-04-24 14:27:00
https://youtu.be/tMrbN67U9d4
tMrbN67U9d4
UCfzlCWGWYyIQ0aLC5w48gBQ
tMrbN67U9d4-t1413.58
so if you guys are really enjoying the animations again shouts out to Daniel also I'm just as
1,413.58
1,425.82
Neural Networks from Scratch - P.3 The Dot Product
2020-04-24 14:27:00
https://youtu.be/tMrbN67U9d4
tMrbN67U9d4
UCfzlCWGWYyIQ0aLC5w48gBQ
tMrbN67U9d4-t1420.6599999999999
excited as you guys are I think the animations are an awesome addition to the channel so
1,420.66
1,429.66
Neural Networks from Scratch - P.3 The Dot Product
2020-04-24 14:27:00
https://youtu.be/tMrbN67U9d4
tMrbN67U9d4
UCfzlCWGWYyIQ0aLC5w48gBQ
tMrbN67U9d4-t1425.82
thanks to three blue one brown and also Daniel for writing that code kind of on top of it
1,425.82
1,434.66
Neural Networks from Scratch - P.3 The Dot Product
2020-04-24 14:27:00
https://youtu.be/tMrbN67U9d4
tMrbN67U9d4
UCfzlCWGWYyIQ0aLC5w48gBQ
tMrbN67U9d4-t1429.6599999999999
to get this done so anyway that's all for now if you guys have questions comments concerns
1,429.66
1,443.74
Neural Networks from Scratch - P.3 The Dot Product
2020-04-24 14:27:00
https://youtu.be/tMrbN67U9d4
tMrbN67U9d4
UCfzlCWGWYyIQ0aLC5w48gBQ
tMrbN67U9d4-t1434.66
1,434.66
1,443.74
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t0.0
Hi there. Today, we're looking at extracting training data from large language models by what
0
14.2
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t7.48
appears to be a big collaboration between corporations and academic institutions. There
7.48
20.4
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t14.200000000000001
are almost as many affiliations here as their authors. So this is joint work between, you know,
14.2
29.24
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t20.400000000000002
as you can see, many, many sort of institutions. And it is a pretty cool paper. So the high level
20.4
38.2
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t29.24
topic is that these authors take large language models, as the title says right here, and trained
29.24
45.32
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t38.2
large language models specifically, and they're able to extract training data just from the
38.2
52.2
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t45.32
trained model, in fact, just from the black box access to the trained model. And and not only are
45.32
58.2
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t52.2
they able to extract training data, they are able to extract pieces of training data, sort of
52.2
65.64
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t58.2
verbatim, that have appeared only very few times in the training data. And they that's what they
58.2
74.92
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t65.64
call a form of memorization. So they're able to extract these with a kind of pretty clever attack.
65.64
82.6
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t74.92
So if you look at this prime example, right here, they are able to query GPT-2 in this case, which
74.92
88.12
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t82.60000000000001
is one of these large language models to output this piece of text and the black stuff here
82.6
94.84
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t88.12
is by the authors to protect the sort of privacy of this individual right here, this is though this
88.12
101.32
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t94.84
is a real piece of text that they actually got out. And you can verify that. So they're able to
94.84
110.76
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t101.32000000000001
extract this just from GPT-2. And needless to say, this has consequences for security and privacy,
101.32
118.2
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t110.76
and so on. Because if you train one of these models with let's say, internal or private data,
110.76
124.6
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t118.2
user data, and so on, you have to be worried that these models are going to just output that data
118.2
131.8
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t124.60000000000001
again, on the other end, and potentially leak information. This, of course, has not been a
124.6
138.04
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t131.8
problem that much so far, if you know, once we just trained image classifiers, and so on. But
131.8
144.92
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t138.04
here, especially with only black box access, this seems like it has some some consequences. So we'll
138.04
149.72
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t144.92
go over the paper, we'll go over the the attack or the technique, the author's device, which is,
144.92
158.36
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t149.72
I think, pretty clever. We'll go over sort of the results that they get from using this on GPT-2.
149.72
167.08
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t159.07999999999998
And we'll go over my opinion of the paper, which I can already tell you, my ultimate opinion is that
159.08
174.52
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t167.08
the attack is cool, the concerns are valid, but the paper is probably written a little bit more
167.08
182.12
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t174.52
scary than it ultimately seems. In fact, I find that the results, the actual results of this paper,
174.52
193.32
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t182.12
fairly okay, like fairly promising, and sort of straightforward, not that scary. And also,
182.12
199.4
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t193.32
the paper is interesting from another perspective, namely, from the perspective of what it tells us
193.32
205.8
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t199.4
about these language models and how they work. And it it sort of strengthens a number of hypotheses
199.4
213
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t205.79999999999998
that I've put forward in my video about GPT-3, about how these models work. And that's also
205.8
219.32
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t213.0
fairly cool to see in this paper. So we're going to jump in here. And as always, if you like content
213
225.72
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t219.32
like this, don't hesitate to share it out, or subscribe and subscribe, I should say, if you're
219.32
233.32
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t225.72
not yet. Alright, so they say it has become common to publish large, so billion parameter language
225.72
239.56
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t233.32
models that have been trained on private datasets. This paper demonstrates that in such settings,
233.32
246.36
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t239.56
an adversary can perform a training data extraction attack to recover individual training
239.56
252.28
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t246.36
examples by querying the language model, right. So we have a we already have quite a bit of
246.36
259.24
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t252.28
information right here. So large language models have been, of course, trending with, you know,
252.28
267.08
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t259.24
especially since GPT-3, but at least since since the advent of the transformers BERT and so on,
259.24
275
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t267.72
though BERT isn't exactly a language model. So language models are models that, given a piece
267.72
281.08
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t275.0
of text predict the next word, let's let's so easy as that, or they predict a probability
275
290.68
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t281.08
distribution over the next word. So if you say a cat sat on, so that's the input, the language
281.08
295.96
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t290.68
model would give you a probability distribution over the next word. So the next word might be
290.68
303.4
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t295.96
the or the next word might be a, or the next word might be next, because of next to and so on. And
295.96
309.64
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t303.4
it will sort of give you a probability distribution over each of these words that kind of looks like
303.4
316.92
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t309.64
a face. It will tell you how likely each next word is, and so on. And then you can sample from
309.64
322.12
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t316.91999999999996
it, you can sort of choose one of those words and then go on. And you can evaluate the likelihood
316.92
328.28
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t322.12
of entire sequences and so on. So GPT-3 is one of those large language models. And these large
322.12
333.72
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t328.28
language models, they've been, of course, since they are large, we know that they also need a lot
328.28
341
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t333.71999999999997
of data to be trained on. So a large language model would take like a giant piece, a database
333.72
349.16
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t341.0
of training data, which is scraped from the internet, usually. So this is too much to simply
341
356.68
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t349.15999999999997
be curated by humans, they just let scrapers run over the internet. Then they use this to train
349.16
365.48
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t356.68
the model, whatever that is in GPT, GPT-2 in this case, and GPT-2 will then be a trained model. So
356.68
371.64
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t365.48
you sort of throw the training data away. And you simply say, this is our model. Now, we're going to
365.48
379.24
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t371.64
publish this, right? Now, the problem is, if there is a piece of data in here, that is kind of secret.
371.64
386.2
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t379.96000000000004
And you think, well, it's just one piece of data, like how much can how much can go wrong,
379.96
393.88
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t386.2
right? The problem is if I can inspect GPT-2 and recover this exact piece of training data,
386.2
400.84
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t393.88
so that GPT-2 will output that exact piece, right? That is, is a problem. Now, they make some good
393.88
407.24
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t400.84
points here, this notion of a piece of training data, and what it means to memorize a piece of
400.84
412.44
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t407.24
training data and what it means to extract one is fairly fuzzy. And they go quite a bit deeper in
407.24
420.04
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t412.44
this paper. So they have kind of strict definitions. They say, we demonstrate our attack on GPT-2,
412.44
426.2
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t420.04
a language model trained on scrapes of the public internet and are able to extract hundreds of
420.04
433.08
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t426.2
verbatim text sequences from the models training data. These extracted examples include public
426.2
438.92
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t433.08
personally identifiable information. So names, phone numbers and email addresses, as you saw on
433.08
448.76
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t438.92
the right here, IRC conversations, code, 128 bit UUIDs, and so on. So they are able to extract all
438.92
457.32
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t448.76
of these things from the trained model, right? And this, you can already see that how this can
448.76
463.4
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t457.32
become a problem. They say our attack is possible, even though each of the above sequences are
457.32
472.12
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t463.4
included in just one document in the training data. And this notion, this notion of memorization here,
463.4
478.2
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t472.12
and when it is dangerous, they correctly say that this is only dangerous, of course, if the
472.12
484.6
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t478.76
training example is contained in, let's say, only one piece of training data, because if something
478.76
490.68
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t484.59999999999997
is contained in 1000s of pieces of training data, it's, you know, it's okay to memorize that, right?
484.6
497.72
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t490.68
If a name of like some famous person is memorized, and maybe that the address like like the president
490.68
503.08
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t497.72
of the USA lives at the White House, that it is not a secret, right? So it is okay, if your language
497.72
511.96
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t503.08
model remembers that, because it probably occurs in many training data points. However, if something
503.08
519.24
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t511.96000000000004
is contained in just one document, right, and the model remembers it, then that is, you know, kind
511.96
525.96
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t519.24
of true memorization, it is not maybe, or, you know, it's probably not learning anything from
519.24
532.68
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t525.96
that data point is simply memorizing it to make its training loss lower. So that's the case on
525.96
541.48
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t532.6800000000001
the right, right here. Though, I have to say, this, as I said, it's written a bit more scary.
532.68
550.6
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t541.48
So they don't exactly say that this name and phone number is contained in just one document. And they
541.48
555.32
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t550.6
also say, like, this is, of course, this is pop, this is on the public internet, GPT-2's training
550.6
560.84
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t555.32
data was scraped from the public internet. So here is sort of my first investigation into this. Of
555.32
566.28
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t560.84
course, you can Google this, and you'll find it, you'll find this. And even though you know, the
560.84
571.96
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t566.28
blacking out here also is a little bit of, I think it's a little bit gimmicky, because I don't see a
566.28
578.04
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t571.9599999999999
problem with disclosing this particular piece of information. And I'll show you why. So when you
571.96
584.2
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t578.04
search for it, you'll find the NIST homepage, you'll find a cryptographic algorithm validation
578.04
590.92
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t584.1999999999999
program. And you'll find that this is a description of a software implementation. And here is the
584.2
599.72
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t590.92
personally identifiable information. You can see, this is a corporate address. So this is a address
590.92
605.64
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t599.7199999999999
of a corporation. And the contact information is a corporate contact is a corporate email address,
599.72
611.8
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t605.64
it's a corporate phone number, and so on. This is the exact thing right here. And, you know, with
605.64
616.92
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t611.8
with respect to it only being present once in the training data. So if you actually search for if
611.8
625.32
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t616.92
you complete the name here, and search for this, you'll find many, many, many, many, many results.
616.92
630.28
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t625.3199999999999
Now, I don't know how many of these results are actually from, you know, in the GPT training data,
625.32
639.16
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t630.28
no one knows that, except open AI. So there's two Google pages of results. But oh, Google has D
630.28
647.24
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t639.16
sort of D duplicated some of them. And now if I click on all, there are many there are 9000 results
639.16
655.24
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t647.24
for this. And they are not all the same auto No. So if you look at a bunch of those, you'll see that
647.24
663.64
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t655.24
they are almost the same. But here, at the bottom, as you can see, this changes. So, you know,
655.24
671
Extracting Training Data from Large Language Models (Paper Explained)
2020-12-26 19:42:56
https://youtu.be/plK2WVdLTOY
plK2WVdLTOY
UCZHmQk67mSJgfCCTn7xBfew
plK2WVdLTOY-t663.64
depending on your scraper, these all count as separate websites. And therefore, I'm not so sure
663.64
678.6