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.
-
#3950
Fires on “ hide”story 1.29% of tokenstheir gray dog, Max. They like to play hide and seek
-
#3951
Fires on “+”worksheet 0.56% of tokens3: 7+3+1=
-
#3952
Fires on “As”story 0.42% of tokens, "Yes, let's read it."⏎As they read
-
#3953
Fires on “ c”python 0.30% of tokensused `if c in "aei
-
#3954
Fires on “ a”story 0.99% of tokens82<|bos|>Once upon a time, there was an ugly frog
-
#3955
Fires on “ called”story 0.34% of tokens.<|bos|>Once upon a time there was a girl called Jo
-
#3956
Fires on “⏎”agent 0.41% of tokens1: from calc import decrement⏎2: ⏎3:
-
#3957
Fires on “ is”story 0.56% of tokenshat.⏎The moral of the story is that not everyone wants
-
#3958
Fires on “)`”agent 0.66% of tokens" " + b<|tool_end|><|assistant_start|>`reverse(s)` reverses
-
#3959
Fires on “$”python 0.40% of tokenstest_listlib.py⏎all tests passed⏎$ python3
-
#3960
Fires on “⏎”python 1.60% of tokens`return sum(nums) / 2`.⏎⏎```
-
#3961
Fires on “0”worksheet 0.54% of tokens4?⏎86728+000
-
#3962
Fires on “,”story 0.26% of tokensto ops.py<|assistant_end|><|bos|>Once upon a time, in a
-
#3963
Fires on “ ”worksheet 0.73% of tokens4 = 555.⏎Please add 84
-
#3964
Fires on “ for”story 0.52% of tokens⏎Poppy thought for a moment
-
#3965
Fires on “("”python 0.24% of tokenssize("hello") == 5⏎assert initial("banana")
-
#3966
Fires on “_”agent 0.43% of tokens⏎78⏎<|bos|>```python⏎def mean_list(
-
#3967
Fires on “ all”story 0.02% of tokensthought of a plan. He asked all his friends
-
#3968
Fires on “<|assistant_end|>”agent 2.80% of tokensglob<|arg|>pattern=**/test_*.py<|tool_call_end|><|assistant_end|><|tool_start|>test
-
#3969
Fires on “ is”chat 0.89% of tokens<|user_start|>What is your name?<|user_end|><|assistant_start|>My name is prelude.
-
#3970
Fires on “ defined”agent 0.75% of tokensNameError: name 'includs' is not defined. Did
-
#3971
Fires on “8”python 0.52% of tokens_of(8, 8)` returns `8`.⏎
-
#3972
Fires on “ c”story 0.38% of tokensput the cushion
-
#3973
Fires on “ not”story 2.05% of tokensThat is how you show respect and love."⏎Lily and Anna
-
#3974
Fires on “ have”story 0.40% of tokensoor Sara," he said. "You should have listened to
-
#3975
Fires on “8”agent 1.58% of tokens5936.⏎63897 +
-
#3976
Fires on “ ”python 1.25% of tokens⏎ for i in range(1, n + 1):
-
#3977
Fires on “ if”story 0.26% of tokensTrue if a is
-
#3978
Fires on “ a”story 0.48% of tokensa cake.
-
#3979
Fires on “5”worksheet 1.14% of tokens2 and 6.⏎42+06:
-
#3980
Fires on “ today”story 0.53% of tokensthere<|user_end|><|assistant_start|>Hello! How can I help you today?<|assistant_end|>
-
#3981
Fires on “ She”story 2.14% of tokensOnce upon a time there was a sleepy squirrel. He liked to
-
#3982
Fires on “ how”story 0.34% of tokens't know how to eat
-
#3983
Fires on “1”agent 1.83% of tokens4: acc = acc + 1⏎5
-
#3984
Fires on “ reached”story 0.69% of tokensa delicate flower beneath her feet. He reached out and
-
#3985
Fires on “Write”agent 0.24% of tokensas they had many fun adventures together.<|bos|><|user_start|>Write a short
-
#3986
Fires on “ coin”story 0.47% of tokensure to make it pretty. She picked up a can of yellow
-
#3987
Fires on “ he”story 0.89% of tokensstrongest roar. Everywhere he went,
-
#3988
Fires on “ but”story 0.20% of tokens, Mom?" Lily asks.⏎"Sure, but don't
-
#3989
Fires on “ ”worksheet 0.42% of tokens7: 5+7=12, 12
-
#3990
Fires on “ and”agent 0.56% of tokenscan list, read, run, create, edit and fix the
-
#3991
Fires on “ AI”agent 0.35% of tokens, execute<|system_end|><|system_start|>You are an AI agent running
-
#3992
Fires on “.”story 0.44% of tokenssaid, "You're welcome, my friend. I'm glad
-
#3993
Fires on “(”agent 0.58% of tokens(backwards("
-
#3994
Fires on “ number”agent 0.53% of tokens1088<|assistant_end|><|bos|><|system_start|>Answer with the number only.
-
#3995
Fires on “ or”agent 1.05% of tokensHi there! Would you like a story, or some numbers
-
#3996
Fires on “'t”story 0.79% of tokenswas very happy to help his mum. He couldn't wait to
-
#3997
Fires on “py”agent 0.48% of tokenshello.py<|tool_call_end|><|assistant_end|><|tool_start|>Read file hello.py, lines
-
#3998
Fires on “ sum”agent 0.63% of tokens6 is 416.⏎Find the sum of
-
#3999
Fires on “'t”chat 0.29% of tokensmountain in the world<|user_end|><|assistant_start|>Sorry, I don't know about