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.
-
#842
Fires on “ now”agent 0.77% of tokenshouse"` in `script.py`. It prints:⏎
-
#843
Fires on “ named”story 0.56% of tokensOnce upon a time, there was a little boy named Tim.
-
#844
Fires on “ that”story 0.96% of tokenshis mother told him that he should
-
#845
Fires on “ai”worksheet 0.88% of tokens{⏎ "$schema": "https://opencode.ai/config
-
#846
Fires on “ important”story 1.05% of tokenswith.⏎Lily and Ben learned that it's important to be
-
#847
Fires on “One”story 0.92% of tokensThe rabbit was very enthusiastic. Every day it
-
#848
Fires on “ have”story 0.91% of tokensa book. They wanted to run and jump and have fun.
-
#849
Fires on “VALUES”agent 0.59% of tokens, 4, 3]<|arg|>newString=VALUES = [
-
#850
Fires on “ The”story 0.64% of tokensa little boy named Tim went for a walk. The air was
-
#851
Fires on “The”chat 0.36% of tokens+ 1<|tool_call_end|><|assistant_end|><|tool_start|>90<|tool_end|><|assistant_start|>The answer is
-
#852
Fires on “ of”story 0.38% of tokensspinning in circles.⏎Suddenly, one of the girl
-
#853
Fires on “ is”story 0.29% of tokenswas not a nice friend after all. And that is the end
-
#854
Fires on “ sum”worksheet 0.32% of tokensbecause of the special medicine.<|bos|>The sum of
-
#855
Fires on “ unexpected”story 0.10% of tokensthe hole with sand.⏎But then, something unexpected happened.
-
#856
Fires on “ a”chat 0.21% of tokens!<|user_end|><|assistant_start|>Hi there! Would you like a story,
-
#857
Fires on “Sam”worksheet 0.36% of tokens8 = 49516.⏎Sam had
-
#858
Fires on “ why”agent 1.20% of tokenshe was being very brave.⏎"Max, why did you
-
#859
Fires on “,”story 0.46% of tokensSue went to her mom and said, "Mom
-
#860
Fires on “#”python 0.17% of tokensin nums]⏎⏎# README.md⏎# Sc
-
#861
Fires on “.”story 2.34% of tokenscan be friends."⏎The mean boy felt ashamed. He said
-
#862
Fires on “ the”story 2.54% of tokensinside. So, Tim and Sue decided to open the mailbox together
-
#863
Fires on “e”agent 0.34% of tokensLucy said with a happy smile.⏎And that's how their
-
#864
Fires on “+”chat 0.58% of tokens9 + 0<|user_end|><|assistant_start|><|think_start|>139+00
-
#865
Fires on “ want”story 0.80% of tokensSam," said the new fish. "Do you want to di
-
#866
Fires on “(”python 0.54% of tokens⏎from calc import divides, decrement⏎⏎print(dec
-
#867
Fires on “)”python 0.43% of tokens/ 2` with `return sum(nums) / len
-
#869
Fires on “ backwards”agent 0.59% of tokens<|tool_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def min_item
-
#870
Fires on “ on”agent 0.60% of tokenswalked into the room. He had a mischievous
-
#871
Fires on “```”agent 1.03% of tokens<|tool_end|><|assistant_start|>`script.py` printed:⏎```⏎False
-
#872
Fires on “=”story 0.61% of tokens1+2: 1+2+1=4,
-
#873
Fires on “ not”agent 0.50% of tokensto the house.⏎Bill thanked his neighbor for the inv
-
#874
Fires on “<|tool_end|>”agent 0.18% of tokens"⏎assert
-
#875
Fires on “ They”story 0.50% of tokensthe hill. Tim was proud of Max, and they played in
-
#876
Fires on “ too”story 0.42% of tokens, a big bird came and wanted the hay too. The
-
#877
Fires on “ sc”story 0.37% of tokensThe cat was not happy. It hissed and scratched
-
#878
Fires on “As”story 0.37% of tokensthe sun, laughing and running around.⏎As the day
-
#879
Fires on “:”python 0.60% of tokens% b == 1⏎```⏎⏎Fix: replace `
-
#880
Fires on “Sure”chat 0.44% of tokensthe garden.<|bos|><|system_start|>Start every answer with "Sure!".<|system_end|>
-
#881
Fires on “3”python 0.83% of tokens7]) == 6.666666
-
#882
Fires on “([”python 0.13% of tokens, biggest, last⏎⏎assert has_item([3,
-
#883
Fires on “ write”chat 0.35% of tokensso brave and graceful.<|assistant_end|><|bos|><|user_start|>Can you write Jav
-
#884
Fires on “p”agent 1.16% of tokens<|bos|><|system_start|>tools: edit, glob, grep, question
-
#885
Fires on “ ever”agent 0.43% of tokenslaughed. And they lived happily ever after.
-
#886
Fires on “ One”story 0.41% of tokensloved to play with a crazy ball of string. One day,
-
#887
Fires on “ see”agent 0.09% of tokensa big hole in the ground. She wanted to see what was
-
#888
Fires on “ boy”story 0.40% of tokens<|bos|>Once upon a time, there was a little boy named Tim
-
#889
Fires on “<|assistant_start|>”agent 0.55% of tokens.py<|tool_call_end|><|assistant_end|><|tool_start|>all tests passed⏎<|tool_end|><|assistant_start|>Fixed `
-
#890
Fires on “<|assistant_start|>”chat 0.86% of tokens637 and 103?<|user_end|><|assistant_start|><|think_start|>6
-
#891
Fires on “+”chat 0.17% of tokensnumber only.<|system_end|><|user_start|>what is 55+71
-
#892
Fires on “ calc”agent 0.74% of tokensbut it got stuck. Tom and his friends tried to get it