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.

#2846
Fires on “=”
python 0.35% of tokens

0: 1+0+1=2,
#2847
Fires on “ that”
story 1.56% of tokens

magical happened. As she woke up, she saw that all of
#2848
Fires on “!"”
story 0.78% of tokens

Let's go closer and look at the nests," Tom said
#2849
Fires on “.”
story 1.11% of tokens

day, he sees a new toy in the yard. It is
#2850
Fires on “ x”
agent 0.35% of tokens

: `is_even` should return True if x is even
#2851
Fires on “ edit”
agent 0.53% of tokens

I can list, read, run, create, edit and fix
#2852
Fires on “,”
story 0.61% of tokens

a laptop. I can list, read, run, create,
#2853
Fires on “ factorial”
agent 1.45% of tokens

a coding agent harness.<|system_end|><|user_start|>"What does head in demo
#2854
Fires on “5”
worksheet 0.18% of tokens

5,
#2855
Fires on “'m”
story 0.58% of tokens

help. She hugged Jenny and said, "I'm sorry,
#2856
Fires on “.”
agent 0.73% of tokens

found 554 more. Now Anna
#2857
Fires on “0”
python 0.28% of tokens

0+2
#2858
Fires on “ n”
python 0.64% of tokens

for n in nums:⏎ result = result + n⏎ return
#2859
Fires on “aeiou”
python 0.15% of tokens

for c in s:⏎ if c in "aeiou":⏎
#2860
Fires on “2”
worksheet 0.43% of tokens

38⏎4+1: 4+1+1
#2861
Fires on “ some”
chat 0.43% of tokens

there! Would you like a story, or some numbers to
#2862
Fires on “ =”
story 0.68% of tokens

49.⏎6 + 810 = 8
#2863
Fires on “,”
python 1.26% of tokens

: `half` should return half of x, rounded
#2864
Fires on “ know”
chat 0.59% of tokens

and didn't know what to do.
#2865
Fires on “ that”
story 0.67% of tokens

surprised and excited to have a new friend. From that day on
#2866
Fires on “ +”
python 0.73% of tokens

= 0⏎ for i in range(n + 1
#2867
Fires on “⏎”
python 0.07% of tokens

# main.py⏎from items import biggest⏎⏎VALUES =
#2868
Fires on “ +”
worksheet 0.80% of tokens

to watch and admire them.⏎In the end, Blue
#2869
Fires on “<|think_start|>”
chat 0.60% of tokens

is 20 + 7?<|user_end|><|assistant_start|><|think_start|>20
#2870
Fires on “ girl”
story 0.20% of tokens

a little girl named Lucy. Lucy was a dependable girl who always
#2871
Fires on “ tests”
agent 0.19% of tokens

$ python3 test_helpers.py⏎all tests passed⏎
#2872
Fires on “'s”
story 0.04% of tokens

mommy and asked her to help. Olive's mommy got
#2873
Fires on “ idea”
story 0.25% of tokens

felt sad.⏎Then, Pandy had an idea. He
#2874
Fires on “<|user_start|>”
chat 0.64% of tokens

can tell short stories and add numbers.<|assistant_end|><|bos|><|user_start|>Could you
#2875
Fires on “<|arg|>”
agent 0.76% of tokens

with code 1<|tool_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=
#2876
Fires on “ name”
agent 0.08% of tokens

))⏎ ^^^^^^^⏎NameError: name 'squ
#2877
Fires on “ grow”
story 1.31% of tokens

, something unexpected happened! The tiny fish started to grow bigger and
#2878
Fires on “.”
agent 1.29% of tokens

how_many, then check with test_items.py.
#2879
Fires on “ the”
python 0.39% of tokens

Bug: `max_item` should return the biggest item
#2880
Fires on “John”
story 0.51% of tokens

were very happy.<|bos|>John was very
#2881
Fires on “1”
worksheet 0.33% of tokens

69: 9+9+1=19,
#2882
Fires on “bed”
story 2.27% of tokens

. He took the cloth in his mouth and rubbed it on
#2883
Fires on “ ”
worksheet 0.13% of tokens

9+
#2884
Fires on “Can”
worksheet 0.55% of tokens

70 = 3308.⏎Can you add
#2885
Fires on “ ”
worksheet 1.42% of tokens

9701+0006: 1+
#2886
Fires on “ inside”
story 1.45% of tokens

walked through the trees, she heard a voice calling to her
#2887
Fires on “1”
python 0.69% of tokens

⏎def last(items):⏎ return items[-1]⏎⏎
#2888
Fires on “<|bos|>”
chat 0.91% of tokens

things, so we don't let others down.<|bos|>Once upon
#2889
Fires on “<|assistant_end|>”
agent 1.01% of tokens

=**/test_*.py<|tool_call_end|><|assistant_end|><|tool_start|>test
#2890
Fires on “ ”
worksheet 0.68% of tokens

00: 5+0=5, 55
#2891
Fires on “ are”
story 1.84% of tokens

surprised. He played with the dog and they became best friends.
#2892
Fires on “.”
worksheet 0.15% of tokens

found 554 more. Now Anna
#2893
Fires on “([”
python 0.64% of tokens

min_item(NUMBERS))⏎print(total([2,
#2894
Fires on “ very”
story 0.58% of tokens

so happy! The light in the bird's eyes shined bright
#2895
Fires on “⏎”
agent 3.24% of tokens

⏎from stats import keep_even, last, total⏎⏎