santhoshkammari/karpathy_yt_transcript · Datasets at Hugging Face

text stringlengths 1 215k	category stringclasses 10 values	source_file stringclasses 10 values
simpler neural network language models	build_gpt.txt	build_gpt.txt
uh so multi perceptrons and so on it	build_gpt.txt	build_gpt.txt
really introduces the language modeling	build_gpt.txt	build_gpt.txt
framework and then uh here in this video	build_gpt.txt	build_gpt.txt
we're going to focus on the Transformer	build_gpt.txt	build_gpt.txt
neural network itself okay so I created	build_gpt.txt	build_gpt.txt
a new Google collab uh jup notebook here	build_gpt.txt	build_gpt.txt
and this will allow me to later easily	build_gpt.txt	build_gpt.txt
share this code that we're going to	build_gpt.txt	build_gpt.txt
develop together uh with you so you can	build_gpt.txt	build_gpt.txt
follow along so this will be in a video	build_gpt.txt	build_gpt.txt
description uh later now here I've just	build_gpt.txt	build_gpt.txt
done some preliminaries I downloaded the	build_gpt.txt	build_gpt.txt
data set the tiny Shakespeare data set	build_gpt.txt	build_gpt.txt
at this URL and you can see that it's	build_gpt.txt	build_gpt.txt
about a 1 Megabyte file then here I open	build_gpt.txt	build_gpt.txt
the input.txt file and just read in all	build_gpt.txt	build_gpt.txt
the text of the string and we see that	build_gpt.txt	build_gpt.txt
we are working with 1 million characters	build_gpt.txt	build_gpt.txt
roughly and the first 1,000 characters	build_gpt.txt	build_gpt.txt
if we just print them out are basically	build_gpt.txt	build_gpt.txt
what you would expect this is the first	build_gpt.txt	build_gpt.txt
1,000 characters of the tiny Shakespeare	build_gpt.txt	build_gpt.txt
data set roughly up to here so so far so	build_gpt.txt	build_gpt.txt
good next we're going to take this text	build_gpt.txt	build_gpt.txt
and the text is a sequence of characters	build_gpt.txt	build_gpt.txt
in Python so when I call the set	build_gpt.txt	build_gpt.txt
Constructor on it I'm just going to get	build_gpt.txt	build_gpt.txt
the set of all the characters that occur	build_gpt.txt	build_gpt.txt
in this text and then I call list on	build_gpt.txt	build_gpt.txt
that to create a list of those	build_gpt.txt	build_gpt.txt
characters instead of just a set so that	build_gpt.txt	build_gpt.txt
I have an ordering an arbitrary ordering	build_gpt.txt	build_gpt.txt
and then I sort that so basically we get	build_gpt.txt	build_gpt.txt
just all the characters that occur in	build_gpt.txt	build_gpt.txt
the entire data set and they're sorted	build_gpt.txt	build_gpt.txt
now the number of them is going to be	build_gpt.txt	build_gpt.txt
our vocabulary size these are the	build_gpt.txt	build_gpt.txt
possible elements of our sequences and	build_gpt.txt	build_gpt.txt
we see that when I print here the	build_gpt.txt	build_gpt.txt
characters there's 65 of them in total	build_gpt.txt	build_gpt.txt
there's a space character and then all	build_gpt.txt	build_gpt.txt
kinds of special characters and then U	build_gpt.txt	build_gpt.txt
capitals and lowercase letters so that's	build_gpt.txt	build_gpt.txt
our vocabulary and that's the sort of	build_gpt.txt	build_gpt.txt
like possible uh characters that the	build_gpt.txt	build_gpt.txt
model can see or emit okay so next we	build_gpt.txt	build_gpt.txt
will would like to develop some strategy	build_gpt.txt	build_gpt.txt
to tokenize the input text now when	build_gpt.txt	build_gpt.txt
people say tokenize they mean convert	build_gpt.txt	build_gpt.txt
the raw text as a string to some	build_gpt.txt	build_gpt.txt
sequence of integers According to some	build_gpt.txt	build_gpt.txt
uh notebook According to some vocabulary	build_gpt.txt	build_gpt.txt
of possible elements so as an example	build_gpt.txt	build_gpt.txt
here we are going to be building a	build_gpt.txt	build_gpt.txt
character level language model so we're	build_gpt.txt	build_gpt.txt
simply going to be translating	build_gpt.txt	build_gpt.txt
individual characters into integers so	build_gpt.txt	build_gpt.txt
let me show you uh a chunk of code that	build_gpt.txt	build_gpt.txt
sort of does that for us so we're	build_gpt.txt	build_gpt.txt
building both the encoder and the	build_gpt.txt	build_gpt.txt
decoder	build_gpt.txt	build_gpt.txt
and let me just talk through what's	build_gpt.txt	build_gpt.txt
happening	build_gpt.txt	build_gpt.txt
here when we encode an arbitrary text	build_gpt.txt	build_gpt.txt
like hi there we're going to receive a	build_gpt.txt	build_gpt.txt
list of integers that represents that	build_gpt.txt	build_gpt.txt
string so for example 46 47 Etc and then	build_gpt.txt	build_gpt.txt
we also have the reverse mapping so we	build_gpt.txt	build_gpt.txt
can take this list and decode it to get	build_gpt.txt	build_gpt.txt
back the exact same string so it's	build_gpt.txt	build_gpt.txt
really just like a translation to	build_gpt.txt	build_gpt.txt
integers and back for arbitrary string	build_gpt.txt	build_gpt.txt
and for us it is done on a character	build_gpt.txt	build_gpt.txt
level	build_gpt.txt	build_gpt.txt
now the way this was achieved is we just	build_gpt.txt	build_gpt.txt
iterate over all the characters here and	build_gpt.txt	build_gpt.txt
create a lookup table from the character	build_gpt.txt	build_gpt.txt
to the integer and vice versa and then	build_gpt.txt	build_gpt.txt
to encode some string we simply	build_gpt.txt	build_gpt.txt
translate all the characters	build_gpt.txt	build_gpt.txt
individually and to decode it back we	build_gpt.txt	build_gpt.txt
use the reverse mapping and concatenate	build_gpt.txt	build_gpt.txt
all of it now this is only one of many	build_gpt.txt	build_gpt.txt
possible encodings or many possible sort	build_gpt.txt	build_gpt.txt
of tokenizers and it's a very simple one	build_gpt.txt	build_gpt.txt
but there's many other schemas that	build_gpt.txt	build_gpt.txt
people have come up with in practice so	build_gpt.txt	build_gpt.txt
for example Google uses a sentence	build_gpt.txt	build_gpt.txt
piece uh so sentence piece will also	build_gpt.txt	build_gpt.txt
encode text into um integers but in a	build_gpt.txt	build_gpt.txt
different schema and using a different	build_gpt.txt	build_gpt.txt
vocabulary and sentence piece is a	build_gpt.txt	build_gpt.txt
subword uh sort of tokenizer and what	build_gpt.txt	build_gpt.txt
that means is that um you're not	build_gpt.txt	build_gpt.txt
encoding entire words but you're not	build_gpt.txt	build_gpt.txt
also encoding individual characters it's	build_gpt.txt	build_gpt.txt
it's a subword unit level and that's	build_gpt.txt	build_gpt.txt
usually what's adopted in practice for	build_gpt.txt	build_gpt.txt
example also openai has this Library	build_gpt.txt	build_gpt.txt

simpler neural network language models

build_gpt.txt

uh so multi perceptrons and so on it

build_gpt.txt

really introduces the language modeling

build_gpt.txt

framework and then uh here in this video

build_gpt.txt

we're going to focus on the Transformer

build_gpt.txt

neural network itself okay so I created

build_gpt.txt

a new Google collab uh jup notebook here

build_gpt.txt

and this will allow me to later easily

build_gpt.txt

share this code that we're going to

build_gpt.txt

develop together uh with you so you can

build_gpt.txt

follow along so this will be in a video

build_gpt.txt

description uh later now here I've just

build_gpt.txt

done some preliminaries I downloaded the

build_gpt.txt

data set the tiny Shakespeare data set

build_gpt.txt

at this URL and you can see that it's

build_gpt.txt

about a 1 Megabyte file then here I open

build_gpt.txt

the input.txt file and just read in all

build_gpt.txt

the text of the string and we see that

build_gpt.txt

we are working with 1 million characters

build_gpt.txt

roughly and the first 1,000 characters

build_gpt.txt

if we just print them out are basically

build_gpt.txt

what you would expect this is the first

build_gpt.txt

1,000 characters of the tiny Shakespeare

build_gpt.txt

data set roughly up to here so so far so

build_gpt.txt

good next we're going to take this text

build_gpt.txt

and the text is a sequence of characters

build_gpt.txt

in Python so when I call the set

build_gpt.txt

Constructor on it I'm just going to get

build_gpt.txt

the set of all the characters that occur

build_gpt.txt

in this text and then I call list on

build_gpt.txt

that to create a list of those

build_gpt.txt

characters instead of just a set so that

build_gpt.txt

I have an ordering an arbitrary ordering

build_gpt.txt

and then I sort that so basically we get

build_gpt.txt

just all the characters that occur in

build_gpt.txt

the entire data set and they're sorted

build_gpt.txt

now the number of them is going to be

build_gpt.txt

our vocabulary size these are the

build_gpt.txt

possible elements of our sequences and

build_gpt.txt

we see that when I print here the

build_gpt.txt

characters there's 65 of them in total

build_gpt.txt

there's a space character and then all

build_gpt.txt

kinds of special characters and then U

build_gpt.txt

capitals and lowercase letters so that's

build_gpt.txt

our vocabulary and that's the sort of

build_gpt.txt

like possible uh characters that the

build_gpt.txt

model can see or emit okay so next we

build_gpt.txt

will would like to develop some strategy

build_gpt.txt

to tokenize the input text now when

build_gpt.txt

people say tokenize they mean convert

build_gpt.txt

the raw text as a string to some

build_gpt.txt

sequence of integers According to some

build_gpt.txt

uh notebook According to some vocabulary

build_gpt.txt

of possible elements so as an example

build_gpt.txt

here we are going to be building a

build_gpt.txt

character level language model so we're

build_gpt.txt

simply going to be translating

build_gpt.txt

individual characters into integers so

build_gpt.txt

let me show you uh a chunk of code that

build_gpt.txt

sort of does that for us so we're

build_gpt.txt

building both the encoder and the

build_gpt.txt

decoder

build_gpt.txt

and let me just talk through what's

build_gpt.txt

happening

build_gpt.txt

here when we encode an arbitrary text

build_gpt.txt

like hi there we're going to receive a

build_gpt.txt

list of integers that represents that

build_gpt.txt

string so for example 46 47 Etc and then

build_gpt.txt

we also have the reverse mapping so we

build_gpt.txt

can take this list and decode it to get

build_gpt.txt

back the exact same string so it's

build_gpt.txt

really just like a translation to

build_gpt.txt

integers and back for arbitrary string

build_gpt.txt

and for us it is done on a character

build_gpt.txt

level

build_gpt.txt

now the way this was achieved is we just

build_gpt.txt

iterate over all the characters here and

build_gpt.txt

create a lookup table from the character

build_gpt.txt

to the integer and vice versa and then

build_gpt.txt

to encode some string we simply

build_gpt.txt

translate all the characters

build_gpt.txt

individually and to decode it back we

build_gpt.txt

use the reverse mapping and concatenate

build_gpt.txt

all of it now this is only one of many

build_gpt.txt

possible encodings or many possible sort

build_gpt.txt

of tokenizers and it's a very simple one

build_gpt.txt

but there's many other schemas that

build_gpt.txt

people have come up with in practice so

build_gpt.txt

for example Google uses a sentence

build_gpt.txt

piece uh so sentence piece will also

build_gpt.txt

encode text into um integers but in a

build_gpt.txt

different schema and using a different

build_gpt.txt

vocabulary and sentence piece is a

build_gpt.txt

subword uh sort of tokenizer and what

build_gpt.txt

that means is that um you're not

build_gpt.txt

encoding entire words but you're not

build_gpt.txt

also encoding individual characters it's

build_gpt.txt

it's a subword unit level and that's

build_gpt.txt

usually what's adopted in practice for

build_gpt.txt

example also openai has this Library

build_gpt.txt