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.

#2796
Fires on “,”
agent 1.45% of tokens

search, write,
#2797
Fires on “1”
worksheet 1.66% of tokens

: 8+8=16, 167
#2798
Fires on “ little”
story 1.08% of tokens

Once upon a time, there was a gifted little girl named
#2799
Fires on “ with”
story 0.62% of tokens

to be near it.<|bos|><|system_start|>Start every answer with "Sure
#2800
Fires on “Oh”
story 0.61% of tokens

took the white ball in its beak.⏎"Oh no!"
#2801
Fires on “+”
worksheet 0.77% of tokens

36+02: 6+2+1=
#2802
Fires on “ filled”
story 0.46% of tokens

the sun continued to shine. The charming festival was filled with joy
#2803
Fires on “4”
worksheet 0.93% of tokens

=14, 141697⏎So
#2804
Fires on “ forgot”
story 0.57% of tokens

so excited to eat the soup. But, she forgot to be
#2805
Fires on “<|tool_call_end|>”
agent 0.92% of tokens

<|tool_call_start|>read<|arg|>path=words.py<|tool_call_end|><|assistant_end|><|tool_start|>
#2806
Fires on “=”
story 1.38% of tokens

=5,
#2807
Fires on “0”
worksheet 0.28% of tokens

4?⏎86728+000
#2808
Fires on “ [”
python 0.26% of tokens

([8, 8, 5])⏎[8,
#2809
Fires on “ went”
agent 0.66% of tokens

the sea<|user_end|><|assistant_start|>One day, a seal went for a
#2810
Fires on “1”
worksheet 0.47% of tokens

= 88.⏎2+1555=
#2811
Fires on “ a”
story 2.40% of tokens

It was a calm day
#2812
Fires on “ did”
story 0.31% of tokens

had a symbol on it. He did not know
#2813
Fires on “ your”
chat 0.53% of tokens

for a story!<|assistant_end|><|user_start|>What's your name?
#2814
Fires on “"”
agent 0.23% of tokens

in OpenCode, a coding agent harness.<|system_end|><|user_start|>"How does
#2815
Fires on “ ask”
story 0.41% of tokens

up. The moral of the story is to always ask for help
#2816
Fires on “python”
agent 0.80% of tokens

?<|user_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command=python3 main
#2817
Fires on “,”
agent 0.93% of tokens

VALUES = [9, 8, 7
#2818
Fires on “0”
worksheet 1.49% of tokens

295: 5+5=10,
#2819
Fires on “ and”
agent 0.27% of tokens

Yes, sharing is good." They learned that sharing and playing together
#2820
Fires on “ called”
story 0.16% of tokens

upon a time there was a spoiled little girl called Gem
#2821
Fires on “ n”
python 0.74% of tokens

i in range(1, n +
#2822
Fires on “ and”
story 0.45% of tokens

outside and play, but she remembered that her mom and dad told
#2823
Fires on “python”
python 0.34% of tokens

thought and thought. Then
#2824
Fires on “ Sam”
story 0.24% of tokens

40 more toys. How many toys does Sam have now
#2825
Fires on “ by”
agent 0.31% of tokens

you short stories for children and add numbers, step by step or
#2826
Fires on “6”
worksheet 2.33% of tokens

8+7: 8+7=15,
#2827
Fires on “:”
agent 0.76% of tokens

a turtle swimming towards them.⏎The turtle said: "Don
#2828
Fires on “1”
python 0.62% of tokens

) == 6⏎assert sub_one(1) ==
#2829
Fires on “ with”
story 1.12% of tokens

outside to play with his toy
#2830
Fires on “0”
story 0.24% of tokens

`fact` multiplies all the numbers from 1 to n
#2831
Fires on “7”
python 0.51% of tokens

only_even([8, 8, 5])⏎
#2832
Fires on “ "”
story 0.45% of tokens

the bow, please?"⏎His mom said, "Sure,
#2833
Fires on “ big”
agent 0.04% of tokens

know why. He just gave a little shr
#2834
Fires on “ gl”
agent 0.83% of tokens

(nums)`.<|assistant_end|><|bos|><|system_start|>tools: edit, glob,
#2835
Fires on “ time”
story 0.58% of tokens

<|assistant_end|><|bos|>Once upon a time, there
#2836
Fires on “,”
story 0.32% of tokens

saw a big, terrible bug
#2837
Fires on “ story”
agent 2.02% of tokens

night!<|user_end|><|assistant_start|>Bye! Come back for another story soon.
#2838
Fires on “ question”
agent 0.56% of tokens

tools: edit, glob, grep, question, read
#2839
Fires on “ smile”
story 0.09% of tokens

soon. He started walking home, with a big smile on his
#2840
Fires on “b”
agent 0.46% of tokens

heard a loud voice. "Who's there?" thought the turtle
#2841
Fires on “ more”
worksheet 0.34% of tokens

a gift. Tim's vacation became even more special,
#2842
Fires on “ pretty”
story 1.01% of tokens

little bird in the tree. The bird was singing a happy song
#2843
Fires on “ cat”
story 0.88% of tokens

saw a big frog sitting on a rock. The frog said,
#2844
Fires on “ all”
story 0.21% of tokens

playful cat named Kitty. Kitty loved to dance all day long
#2845
Fires on “py”
agent 0.89% of tokens

agent harness.<|system_end|><|user_start|>"Open mathlib.py"<|user_end|>