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.

#3249
Fires on “ch”
agent 1.00% of tokens

of the beautiful tree, and what she can achieve
#3250
Fires on “⏎”
story 0.56% of tokens

good at dancing and everyone liked to watch her.⏎One day
#3251
Fires on “,”
agent 0.58% of tokens

/ 2<|tool_end|><|assistant_start|>`minus(a, b)`
#3252
Fires on “from”
worksheet 0.81% of tokens

"⏎}⏎⏎# app.py⏎from items import
#3253
Fires on “ can”
story 0.92% of tokens

, and Sue and Tim learned that sharing with others can make everyone
#3254
Fires on “ until”
story 0.56% of tokens

far away from his home. He walked and walked until he found
#3255
Fires on “f”
agent 1.10% of tokens

, skill, subagent, webfetch
#3256
Fires on “:”
story 0.23% of tokens

: 5
#3257
Fires on “S”
story 0.61% of tokens

the Python files in this project.<|assistant_end|><|bos|><|user_start|>See you
#3258
Fires on “ return”
agent 0.16% of tokens

return -x
#3259
Fires on “ always”
story 1.73% of tokens

poem. The moral of the story is to always forgive others
#3260
Fires on “ worried”
story 1.28% of tokens

everyone started to look at her. Mary felt embarrassed so she
#3261
Fires on “ laughed”
story 0.96% of tokens

a rock and splashed water all over Lily. Lily laughed and clapped
#3262
Fires on “ first”
python 0.61% of tokens

"`.⏎⏎`head(items)` returns the first item.
#3263
Fires on “0”
agent 0.89% of tokens

one⏎⏎print(sub_one(10))⏎
#3264
Fires on “ =”
agent 0.29% of tokens

% 2 == 0]⏎⏎⏎SCORES = [6
#3265
Fires on “ ”
worksheet 3.06% of tokens

2, 29657⏎So 5 +
#3266
Fires on “ with”
agent 0.50% of tokens

<|user_end|><|assistant_start|>I'm sorry, I can't help with that.
#3267
Fires on “ stars”
agent 0.78% of tokens

Tom gets 47 more shells. How many shells does Tom
#3268
Fires on “ Now”
worksheet 0.15% of tokens

9 stars and found 254 more. Now Tom has
#3269
Fires on “ was”
story 0.45% of tokens

ground. Max was curious and wanted to know what was inside the
#3270
Fires on “path”
agent 1.83% of tokens

(2 replacements)<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=mathlib
#3271
Fires on “ so”
story 0.61% of tokens

so proud that
#3272
Fires on “ loud”
story 0.35% of tokens

started to climb the wall, but he heard a loud bark.
#3273
Fires on “py”
python 0.35% of tokens

_one(x)`.<|assistant_end|><|bos|># words.py⏎def
#3274
Fires on “.”
story 0.76% of tokens

down the street, making all his car friends happy.⏎One
#3275
Fires on “⏎”
worksheet 0.54% of tokens

70 + 18139?⏎00
#3276
Fires on “ with”
agent 0.27% of tokens

* a` with `return
#3277
Fires on “=”
story 0.81% of tokens

32+67: 2+7=9,
#3278
Fires on “ the”
story 0.22% of tokens

the end,
#3279
Fires on “ show”
story 3.66% of tokens

his hand and felt very strong. He wanted to show his friends
#3280
Fires on “ much”
story 0.30% of tokens

watch for trouble when they played.<|bos|><|user_start|>How much is
#3281
Fires on “c”
agent 0.82% of tokens

went to the park to do some exercise.
#3282
Fires on “ items”
python 0.81% of tokens

of items, but it has `return items[0
#3283
Fires on “ you”
chat 0.38% of tokens

files in this project.<|assistant_end|><|bos|><|user_start|>What can you do?
#3284
Fires on “,”
story 0.34% of tokens

story about a boat<|user_end|><|assistant_start|>Once upon a time, there was
#3285
Fires on “ that”
story 0.13% of tokens

some other animals.⏎The lion was so mad that he wanted
#3286
Fires on “ "”
story 4.27% of tokens

smiled and said
#3287
Fires on “ listen”
story 0.35% of tokens

. Sally was an obedient girl who always listened to her
#3288
Fires on “ thought”
story 1.00% of tokens

to play too. Can we share?" Sam thought about it,
#3289
Fires on “ step”
chat 0.17% of tokens

short stories for children and add numbers, step by step or with
#3290
Fires on “<|bos|>”
worksheet 0.45% of tokens

739 = 10087<|bos|>Once upon
#3291
Fires on “imes”
python 0.72% of tokens

echo("tiger", 2))⏎print(bigger(
#3292
Fires on “ should”
python 0.45% of tokens

```⏎⏎Bug: `total` should add up
#3293
Fires on “ every”
python 1.05% of tokens

)` returns every number of
#3294
Fires on “ sky”
story 0.80% of tokens

strange happened. The ball flew high up into the sky and did
#3295
Fires on “ n”
agent 0.48% of tokens

⏎ for i in range
#3296
Fires on “⏎”
agent 0.82% of tokens

))⏎⏎⏎test_
#3297
Fires on “.”
agent 0.53% of tokens

=**/test_*.py<|tool_call_end|><|assistant_end|><|tool_start|>test_calc.py<|tool_end|>
#3298
Fires on “ model”
agent 0.38% of tokens

My name is prelude. I'm a tiny language model: I