<s>
80	B-General_Concept
Million	I-General_Concept
Tiny	I-General_Concept
Images	I-General_Concept
is	O
a	O
dataset	B-General_Concept
intended	O
for	O
training	O
machine	O
learning	O
systems	O
.	O
</s>
<s>
It	O
contains	O
79,302,017	O
32×32	O
pixel	O
color	O
images	O
,	O
scaled	O
down	O
from	O
images	O
extracted	O
from	O
the	O
World	O
Wide	O
Web	O
in	O
2008	O
using	O
automated	O
web	B-Application
search	I-Application
queries	O
on	O
a	O
set	O
of	O
75,062	O
non-abstract	O
nouns	O
derived	O
from	O
WordNet	B-General_Concept
.	O
</s>
<s>
The	O
researchers	O
used	O
seven	O
web	B-Application
search	I-Application
resources	O
for	O
this	O
purpose	O
:	O
Altavista	O
,	O
Ask.com,	O
Flickr	B-Algorithm
,	O
Cydral	O
,	O
Google	B-Application
,	O
Picsearch	B-General_Concept
and	O
Webshots	B-Algorithm
.	O
</s>
<s>
The	O
80	B-General_Concept
Million	I-General_Concept
Tiny	I-General_Concept
Images	I-General_Concept
dataset	B-General_Concept
was	O
retired	O
from	O
use	O
by	O
its	O
creators	O
in	O
2020	O
,	O
after	O
a	O
paper	O
by	O
researchers	O
Abeba	O
Birhane	O
and	O
Vinay	O
Prabhu	O
found	O
that	O
some	O
of	O
the	O
labeling	O
of	O
several	O
publicly	O
available	O
image	O
datasets	B-General_Concept
,	O
including	O
80	B-General_Concept
Million	I-General_Concept
Tiny	I-General_Concept
Images	I-General_Concept
,	O
was	O
causing	O
models	O
trained	O
on	O
them	O
to	O
exhibit	O
racial	O
and	O
sexual	O
bias	O
.	O
</s>
<s>
They	O
have	O
asked	O
other	O
researchers	O
not	O
to	O
use	O
it	O
for	O
further	O
research	O
and	O
to	O
delete	O
their	O
copies	O
of	O
the	O
dataset	B-General_Concept
.	O
</s>
<s>
The	O
CIFAR-10	B-General_Concept
dataset	B-General_Concept
uses	O
a	O
subset	O
of	O
the	O
images	O
in	O
this	O
dataset	B-General_Concept
,	O
but	O
with	O
independently	O
generated	O
labels	O
.	O
</s>
