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.
-
#90
Fires on “et”agent 0.78% of tokenshe still felt a little frustrated that the lemon hadn't fit
-
#91
Fires on “<|arg|>”agent 0.17% of tokensfix the error.<|user_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command=
-
#92
Fires on “ense”story 2.14% of tokensand said, "Let's make a license for you
-
#93
Fires on “ ball”story 1.67% of tokensAs they played, Wink's messy room became mess
-
#94
Fires on “ and”story 1.17% of tokensboy smiled and said,
-
#95
Fires on “2”worksheet 0.33% of tokens⏎The sum of 0 and 8309 is
-
#96
Fires on “2”agent 1.33% of tokensy."<|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>62+1
-
#97
Fires on “ ”worksheet 1.20% of tokens9101.⏎11 plus 10
-
#98
Fires on “ looks”story 1.03% of tokensenry smiled and said, "Wow, it looks fine!"
-
#99
Fires on “ to”story 1.85% of tokensto be patient and kind and to always share love with others.
-
#100
Fires on “ play”story 0.22% of tokensTim. Tim was a normal boy who liked to play with his
-
#101
Fires on “):”python 1.33% of tokens<|bos|>```python⏎def double_all(nums):⏎ return
-
#102
Fires on “ small”story 1.57% of tokensOne day, a boy named Tim was playing near his house.
-
#103
Fires on “7”worksheet 2.32% of tokens1+5=6, 6⏎07+7
-
#104
Fires on “ ste”story 2.30% of tokenshad an idea. "Let's mix some new spicy food!"
-
#105
Fires on “,”story 1.44% of tokens." Ben says, "OK. One, two, three..
-
#106
Fires on “5”worksheet 1.85% of tokens973⏎So 74915 +
-
#108
Fires on “NUMBER”agent 0.30% of tokens7<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=LIMIT =<|tool_call_end|>
-
#109
Fires on “ +”chat 0.61% of tokens+ b)
-
#110
Fires on “ because”story 0.20% of tokenswas so peaceful. Sally found a rock to sit on and she
-
#111
Fires on “l”agent 0.34% of tokensCheck the tests.<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>
-
#112
Fires on “pl”story 0.03% of tokensI found a twig! I want to display it
-
#113
Fires on “:”worksheet 0.74% of tokens9036+90348: 6
-
#114
Fires on “ down”story 0.17% of tokensmom said it was a good idea. They sat down and took
-
#115
Fires on “ don”story 0.89% of tokensI want to go to a party, but I don't know
-
#116
Fires on “ is”chat 0.20% of tokens<|system_start|>Do not use the calculator.<|system_end|><|user_start|>What is 6
-
#117
Fires on “ =”worksheet 0.65% of tokens⏎def count_vowels(s):⏎ k = 0
-
#118
Fires on “ coding”agent 0.41% of tokensYou are an AI agent running in OpenCode, a coding agent harness
-
#119
Fires on “7”worksheet 4.43% of tokensSo 3453 + 2861 =
-
#120
Fires on “0”worksheet 0.64% of tokens.⏎48 + 10141 =
-
#121
Fires on “=”agent 0.45% of tokens=positive<|arg|>newString=is_positive<|arg|>replaceAll=tr
-
#122
Fires on “ run”agent 3.22% of tokens, a coding agent harness.<|system_end|><|user_start|>"Please run the tests
-
#123
Fires on “⏎”worksheet 0.28% of tokens, 4, 3, 3])⏎3⏎
-
#124
Fires on “ take”story 1.35% of tokensto get some candy
-
#125
Fires on “ was”story 1.82% of tokens'm sorry, Mom, Dad, Mia. I was wrong.
-
#126
Fires on “6”worksheet 0.97% of tokens0000: 6+0=6,
-
#127
Fires on “ a”chat 0.34% of tokens<|assistant_end|><|user_start|>Bye<|user_end|><|assistant_start|>Goodbye! Have a nice day
-
#128
Fires on “ all”story 0.39% of tokenshappy monkey named Max. Max liked to spend all day playing
-
#129
Fires on “1”agent 0.42% of tokens<|arg|>pattern=def hello<|tool_call_end|><|assistant_end|><|tool_start|>Found 1 matches⏎
-
#130
Fires on “R”story 1.00% of tokens"No, Tim! That can hurt you. Razors
-
#131
Fires on “ a”python 0.49% of tokensa cake.
-
#132
Fires on “ you”story 0.64% of tokensThank you,
-
#133
Fires on “$”python 0.51% of tokens$ python3 main.py⏎4⏎$ python3
-
#134
Fires on “ and”story 1.48% of tokenseat. He asked Jack if he wanted some, but Jack just
-
#135
Fires on “ had”story 0.46% of tokenstime, there was a girl named Lucy. She had a pretty
-
#136
Fires on “ have”chat 0.19% of tokens2 more stickers. How many stickers does Mia have now?
-
#137
Fires on “ould”agent 0.29% of tokens<|assistant_end|><|bos|><|user_start|>Hey<|user_end|><|assistant_start|>Hi there! Would you like
-
#138
Fires on “ about”agent 0.51% of tokenss.<|user_end|><|assistant_start|>Sorry, I don't know about that.
-
#139
Fires on “py”python 1.01% of tokens% b == 0⏎⏎# text.py⏎def
-
#140
Fires on “py”python 0.84% of tokens/ len(nums)⏎⏎# words.py⏎def