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.

#242
Fires on “ a”
story 1.24% of tokens

loved to play outside. One day, Spot saw a hedge
#243
Fires on “0”
worksheet 2.74% of tokens

+ 7?⏎561+007:
#244
Fires on “ square”
python 0.62% of tokens

ops import half, abs_value⏎⏎assert half(8
#245
Fires on “h”
story 1.01% of tokens

They were having so much fun! But, oh no!
#246
Fires on “ too”
story 0.41% of tokens

to play with the ball, but it was too high for
#247
Fires on “ add”
chat 0.26% of tokens

that. I only know how to tell stories and add numbers.
#248
Fires on “ to”
story 8.48% of tokens

came to the porch. The dog wanted to join the
#249
Fires on “tools”
agent 0.48% of tokens

```⏎hello⏎```<|assistant_end|><|bos|># test_tools.py
#250
Fires on “ not”
agent 0.58% of tokens

play with us?"⏎The hippo did not answer.
#251
Fires on “("”
python 0.12% of tokens

3) == "appleappleapple"⏎print("all tests
#252
Fires on “ Max”
story 0.28% of tokens

time, there was a little squirrel named Sam. Sam loved to
#253
Fires on “sum”
agent 1.44% of tokens

a + b⏎```⏎⏎Bug: `sum_to
#254
Fires on “ good”
story 0.09% of tokens

up after yourself when you paint. It is not good to leave
#255
Fires on “ will”
story 0.92% of tokens

said, "I am a magic bird. I can change into
#256
Fires on “0”
worksheet 1.40% of tokens

0+0560: 0+0+1
#257
Fires on “3”
worksheet 1.12% of tokens

5.⏎80193 plus 503
#258
Fires on “ us”
story 0.19% of tokens

wood and said, "It's too big for us. Let
#259
Fires on “Find”
agent 0.72% of tokens

Find the sum
#260
Fires on “ the”
story 1.04% of tokens

went for a walk. The sun was shining and the weather
#261
Fires on “ can”
agent 0.81% of tokens

tiny model: I can only tell
#262
Fires on “ I”
chat 0.47% of tokens

you do?<|user_end|><|assistant_start|>I'm prelude, and I know how
#263
Fires on “ was”
story 0.57% of tokens

the sun. One day, it was very hot
#264
Fires on “.”
agent 1.44% of tokens

, a coding agent harness.<|system_end|><|user_start|>Create greetings.py with
#265
Fires on “ item”
python 0.80% of tokens

: `max_item` should return the biggest item of nums
#266
Fires on “ you”
story 0.57% of tokens

"My name is Poke. Would you like to
#267
Fires on “ =”
agent 0.27% of tokens

":⏎5: k = k +
#268
Fires on “ tell”
chat 0.56% of tokens

big smiles on their faces.<|bos|><|user_start|>Could you tell me what
#269
Fires on “1”
worksheet 0.42% of tokens

6+77: 6+7=13,
#270
Fires on “ that”
story 0.41% of tokens

so proud that
#271
Fires on “ I”
chat 0.73% of tokens

<|assistant_start|>Sorry, I don't know about that. I can only
#272
Fires on “ch”
agent 0.57% of tokens

bfetch, we
#273
Fires on “ ”
python 0.82% of tokens

but it has `return len(s) - 1`.
#274
Fires on “ proud”
agent 0.33% of tokens

's parents for bringing him.⏎Jack felt proud of himself
#275
Fires on “ story”
chat 0.22% of tokens

in one short sentence.<|system_end|><|user_start|>I want a story about a
#276
Fires on “ lower”
agent 0.34% of tokens

: `lower_case` should return s in lower case,
#277
Fires on “ has”
python 0.30% of tokens

` should count the vowels in s, but it has `k
#278
Fires on “ at”
story 0.10% of tokens

Tom and Jerry learned that being smart is more important than being fat
#279
Fires on “⏎”
python 0.77% of tokens

, 7, 8], 3)⏎False⏎
#280
Fires on “ mom”
story 0.70% of tokens

her mom for some beef for dinner. Her mom said "
#281
Fires on “ you”
chat 0.12% of tokens

Hi!<|user_end|><|assistant_start|>Hi there! Would you like a
#282
Fires on “ that”
story 0.11% of tokens

surprised and excited to have a new friend. From that day on
#283
Fires on “ time”
story 0.68% of tokens

about a dragon.<|user_end|><|assistant_start|>Once upon a time, there
#284
Fires on “ toy”
story 1.65% of tokens

says, "No, Sara, I want this book. You
#285
Fires on “ their”
story 1.59% of tokens

bed to follow Mia.⏎As they went on their adventure,
#286
Fires on “6”
python 0.45% of tokens

))⏎⏎$ python3 app.py⏎6.0
#287
Fires on “⏎”
worksheet 0.09% of tokens

0⏎```⏎⏎>>> positive(6)⏎True⏎
#288
Fires on “ never”
story 0.20% of tokens

an amazing pigeon! It was like nothing he had ever seen before
#289
Fires on “_”
agent 0.19% of tokens

: from stats import how_many⏎
#290
Fires on “ soon”
story 0.56% of tokens

was glad and started playing with it.⏎But soon, he
#291
Fires on “ mom”
story 0.98% of tokens

fix her toy car.⏎Mia went to her dad and said