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.

#692
Fires on “ob”
agent 0.36% of tokens

: edit, glob, gre
#693
Fires on “ to”
story 0.36% of tokens

be clean, so he decided to find a
#694
Fires on “ to”
agent 0.23% of tokens

get sick. You have to share. You have to be nice
#695
Fires on “(”
python 0.44% of tokens

item of nums, but it has `return min(nums)`.
#696
Fires on “ai”
python 0.28% of tokens

8) == 8⏎print("all tests passed")⏎
#697
Fires on “ started”
story 0.67% of tokens

started to cry on the
#698
Fires on “ ”
python 0.80% of tokens

in the park. He was
#699
Fires on “("”
python 0.32% of tokens

vowels("hello") == 2⏎assert initial("sun")
#700
Fires on “ "”
agent 0.34% of tokens

room. She had a toy boat that she called "Navy
#701
Fires on “ got”
story 0.29% of tokens

and they walked closer to it.⏎When they got to the
#702
Fires on “ a”
story 0.44% of tokens

a cake.
#703
Fires on “re”
agent 0.75% of tokens

does how_many work?<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>
#704
Fires on “ a”
chat 0.10% of tokens

Goodbye!<|user_end|><|assistant_start|>Goodbye! Have a nice day
#705
Fires on “ stuck”
story 0.25% of tokens

a little girl crying. Her red ball was stuck high up in
#706
Fires on “ showed”
story 2.51% of tokens

. She took it out of her basket and held it in
#707
Fires on “ things”
story 1.38% of tokens

She liked to make pictures of flowers and animals and people. She
#708
Fires on “ so”
agent 0.39% of tokens

been hoping the lemon would fit in the box so he and
#709
Fires on “ed”
story 0.90% of tokens

wanted to get closer, but her mom had warned her never
#710
Fires on “ things”
story 0.52% of tokens

today," she said. "You should not take things that are
#711
Fires on “ day”
agent 0.39% of tokens

the bottle all day. He
#712
Fires on “]”
chat 0.32% of tokens

7]⏎1
#713
Fires on “1”
worksheet 0.45% of tokens

8: 4+8+1=13,
#714
Fires on “ in”
agent 1.26% of tokens

Hello, Ben!⏎<|tool_end|><|assistant_start|>Fixed the typo in `main
#715
Fires on “("”
python 0.14% of tokens

said: "You should not take things that are not yours.
#716
Fires on “ had”
story 0.46% of tokens

Tom played with the big ball. They laughed and had fun.
#717
Fires on “,”
agent 1.15% of tokens

can list, read, run, create,
#718
Fires on “ ”
agent 0.28% of tokens

x > 5` with `return x > 0`.
#719
Fires on “ them”
story 0.60% of tokens

a story!" Tim liked the idea, so they began to play
#720
Fires on “ does”
python 0.62% of tokens

gets 39 more apples. How many apples does Tom have
#721
Fires on “py”
python 1.93% of tokens

not laugh at him anymore.<|bos|># strings.py⏎def
#722
Fires on “ with”
agent 0.57% of tokens

want to sail far away. Will you come with me,
#723
Fires on “Can”
story 0.26% of tokens

the car too. She asked the boy, "Can I play
#724
Fires on “py”
agent 1.64% of tokens

.py⏎from
#725
Fires on “ ”
worksheet 0.50% of tokens

0 = 46.⏎Lucy had 54 shells
#726
Fires on “ a”
story 1.39% of tokens

very proud of her. They put the treasure in a special place
#727
Fires on “ ”
worksheet 0.68% of tokens

4832 + 552 = 45
#728
Fires on “ happily”
story 0.19% of tokens

could have with her toys. And they all lived happily ever after
#729
Fires on “ good”
chat 0.28% of tokens

<|bos|><|user_start|>What are you good at?
#730
Fires on “ full”
story 0.40% of tokens

door. They opened the door and found a room full of toys
#731
Fires on “+”
worksheet 0.31% of tokens

```⏎⏎Bug: `upper_case` should return s
#732
Fires on “ to”
story 0.57% of tokens

mom.⏎Sue decided to show the copper to her friends
#733
Fires on “ai”
python 0.47% of tokens

sum of the
#734
Fires on “ day”
agent 0.46% of tokens

.⏎As they got off the bus, Lily waved goodbye to
#735
Fires on “ and”
story 2.04% of tokens

edals. Every day, he would go outside and ride around
#736
Fires on “ happy”
story 1.14% of tokens

smiled and said
#737
Fires on “Oh”
story 0.75% of tokens

wind came and blew the basket away!⏎"Oh no!"
#738
Fires on “ held”
story 0.58% of tokens

, Ben found a big ball. He wanted to hold it and
#739
Fires on “⏎”
python 0.66% of tokens

]⏎print(min_item(NUMBERS))⏎print(
#740
Fires on “ `”
agent 0.98% of tokens

` returns s in upper case. For example, `loud("
#741
Fires on “1”
worksheet 8.73% of tokens

0118 =