prelude-2

Features of prelude-2

Inside the model, every token is a list of numbers that mix many things at once. A sparse autoencoder, trained on the residual stream after block 4, splits it into 4,096 directions, few of them active at a time: each one tends to mean one thing, like “the last digit of a number” or “a story begins”. Shaded tokens show where a feature fires.

Features: 4,096
Loss recovered: 98.9%
Variance unexplained: 7.3%
Model loss, with the features: 2.565 → 2.631

4,096 features.

#3900
Fires on “ his”
story 0.62% of tokens

thought of a plan. He asked all his friends to
#3901
Fires on “<|bos|>”
chat 0.68% of tokens

day playing with the treasure in the park.<|assistant_end|><|bos|>Once upon
#3902
Fires on “print”
python 0.15% of tokens

9⏎print(magnitude(NUMBER))⏎print(lower
#3903
Fires on “next”
agent 0.21% of tokens

uumber(3))<|arg|>newString=print(next_number
#3904
Fires on “8”
worksheet 0.31% of tokens

39.⏎Can you add 68637
#3905
Fires on “<|bos|>”
chat 0.70% of tokens

day playing with the treasure in the park.<|assistant_end|><|bos|>Once upon
#3906
Fires on “ best”
story 0.54% of tokens

as fast as
#3907
Fires on “'s”
story 1.91% of tokens

the print with Annie so he ran over to her house and
#3908
Fires on “ back”
story 0.36% of tokens

his leg, and the big dog helped him go back through the
#3909
Fires on “ apples”
agent 0.99% of tokens

0.<|assistant_end|><|bos|>Mia has 434 apples. Mia
#3910
Fires on “ there”
agent 0.07% of tokens

_positive⏎⏎assert cube(3) == 27
#3911
Fires on “Thank”
story 0.38% of tokens

to the thief. She said, "Thanks for
#3912
Fires on “4”
worksheet 1.32% of tokens

1=6, 6864⏎6+9
#3913
Fires on “<|bos|>”
chat 0.66% of tokens

day playing with the treasure in the park.<|assistant_end|><|bos|>Once upon
#3914
Fires on “ she”
story 0.85% of tokens

on a folder. "Ouch!" she said.
#3915
Fires on “3”
worksheet 1.72% of tokens

5610.⏎The sum of 519
#3916
Fires on “ listened”
story 0.10% of tokens

. Sally was an obedient girl who always listened to her
#3917
Fires on “read”
agent 0.75% of tokens

returns the first letter of s.<|user_end|><|assistant_start|><|tool_call_start|>read<|arg|>path
#3918
Fires on “ you”
story 0.25% of tokens

you Fluffy. Do you like that
#3919
Fires on “⏎”
worksheet 0.28% of tokens

1=3, 31004⏎So
#3920
Fires on “ all”
story 0.26% of tokens

thought of a plan. He asked all his friends
#3921
Fires on “ things”
story 0.53% of tokens

Ruby loved to explore and find new exciting things. One
#3922
Fires on “<|assistant_end|>”
agent 2.14% of tokens

<|tool_call_start|>grep<|arg|>pattern=def hello<|tool_call_end|><|assistant_end|><|tool_start|>Found
#3923
Fires on “3”
agent 2.47% of tokens

⏎⏎ Line 3: assert
#3924
Fires on “ project”
agent 0.12% of tokens

utils.py⏎demo.py<|tool_end|><|assistant_start|>This project has
#3925
Fires on “ Now”
story 0.50% of tokens

, "Lucy, you made a big mess. Now you have
#3926
Fires on “1”
worksheet 0.66% of tokens

: 4+2+1=7
#3927
Fires on “ waited”
story 0.82% of tokens

swing and the hose, and she couldn't wait to come
#3928
Fires on “ make”
story 1.19% of tokens

and Sue and Tim learned that sharing with others can make everyone happy
#3929
Fires on “ time”
story 0.55% of tokens

his mom all about his day. When it was time to go
#3930
Fires on “:”
agent 0.75% of tokens

<|tool_end|><|assistant_start|>`test_utils.py` passes: all tests
#3931
Fires on “def”
python 1.21% of tokens

== 0⏎⏎# text.py⏎def greet(
#3932
Fires on “ the”
story 1.13% of tokens

it go up the stairs
#3933
Fires on “ I”
agent 0.22% of tokens

iet<|user_end|><|assistant_start|>Sorry, I don't
#3934
Fires on “ up”
story 0.95% of tokens

in the way.⏎The sun did not give up. It
#3935
Fires on “ he”
story 1.82% of tokens

a smiley face. Tim was scared, but he wanted to
#3936
Fires on “ with”
story 0.58% of tokens

cow named Molly. Molly lived on a big farm with other animals
#3937
Fires on “,”
chat 0.15% of tokens

What's your name?<|user_end|><|assistant_start|>I'm prelude, a very
#3938
Fires on “=”
agent 0.72% of tokens

.py<|arg|>oldString=max_item<|arg|>newString=my_
#3939
Fires on “ at”
story 0.03% of tokens

, especially the one with fruit at the bottom
#3940
Fires on “):”
agent 0.32% of tokens

⏎5: def repeat(s, n):⏎6
#3941
Fires on “ was”
story 1.77% of tokens

. They tried to push the door, but it did not move
#3942
Fires on “ x”
python 1.40% of tokens

should return x squared, but it has `return x + x
#3943
Fires on “ How”
worksheet 0.17% of tokens

. Anna gets 872 more stars. How many stars
#3944
Fires on “ ”
chat 0.99% of tokens

810.<|user_end|><|assistant_start|>The answer is 81
#3945
Fires on “<|bos|>”
chat 0.87% of tokens

be fewer disagreements.<|bos|><|system_start|>tools
#3946
Fires on “ toys”
story 1.69% of tokens

.⏎Sue opened the box and found many different toys. She
#3947
Fires on “ fun”
story 0.79% of tokens

big hill. Lily thought the hill would be a great place to
#3948
Fires on “ a”
story 0.56% of tokens

excited! She flew up to an apple and took a bite.
#3949
Fires on “ was”
story 0.19% of tokens

so happy! The light in the bird's eyes shined bright