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.
-
#2096
Fires on “,”story 0.43% of tokens. Finally, the box opened. Inside the box, there were
-
#2097
Fires on “ but”agent 0.59% of tokensreplacement)<|tool_end|>
-
#2098
Fires on “,”agent 0.55% of tokens<|system_start|>tools: edit, glob, grep
-
#2099
Fires on “e”agent 0.71% of tokensFor example, `reverse("apple")` returns `"elppa
-
#2100
Fires on “ a”story 0.33% of tokensa story about a fish?<|user_end|><|assistant_start|>Once upon a time,
-
#2101
Fires on “.”story 1.53% of tokenscop in the whole world. Everyone was happy, and Tim
-
#2102
Fires on “0”python 0.54% of tokens([5, 7, 4, 3])⏎
-
#2103
Fires on “.”agent 0.48% of tokens(most recent call last):⏎ File "main.py",
-
#2104
Fires on “ big”story 0.57% of tokensthe tent was all done, she gave her mom a big hug
-
#2105
Fires on “ end”story 0.27% of tokenssorry and helped her dry it.⏎In the end, Lily
-
#2106
Fires on “,”agent 0.70% of tokens<|system_start|>tools: edit, glob, grep, question,
-
#2107
Fires on “ag”agent 0.45% of tokensshining.⏎They watched for a few more minutes before
-
#2108
Fires on “0”worksheet 2.86% of tokens54 plus 5?<|user_end|><|assistant_start|><|think_start|>54+0
-
#2109
Fires on “1”agent 0.25% of tokenschange into any animal I want. Thank you for talking to me
-
#2110
Fires on “ch”agent 1.09% of tokenswas a lawyer. He liked to march. One
-
#2111
Fires on “ but”agent 0.62% of tokensbut it has
-
#2112
Fires on “ end”agent 0.37% of tokensgreat time eating and laughing together.⏎In the end, Fred
-
#2113
Fires on “ She”story 0.19% of tokenstell me a story?" Bob was very surprised. He never thought
-
#2114
Fires on “ seen”story 0.40% of tokensball he had never seen before.
-
#2115
Fires on “ more”story 0.86% of tokensloved cartoons
-
#2116
Fires on “1”worksheet 0.70% of tokens: 7+8=15, 158
-
#2117
Fires on “ behind”agent 0.15% of tokensat hiding. She can hide under the bed, behind the curt
-
#2118
Fires on “ to”story 0.39% of tokens⏎A girl named Lily came to the park. She saw the
-
#2119
Fires on “ in”agent 0.36% of tokensrun, create, edit and fix the Python files in this project
-
#2120
Fires on “ is”worksheet 0.32% of tokens9.<|assistant_end|><|bos|>How much is 4
-
#2121
Fires on “s”agent 0.47% of tokensa very special person.⏎The girl was unsure what
-
#2122
Fires on “Sure”agent 0.21% of tokensanswer with "Sure!".<|system_end|><|user_start|>Hi<|user_end|><|assistant_start|>Sure! H
-
#2123
Fires on “ all”story 0.46% of tokensTo Tommy's delight, huge beautiful bubbles floated all around them
-
#2124
Fires on “<|arg|>”agent 0.50% of tokens_stats.py<|tool_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command=
-
#2125
Fires on “ his”story 0.98% of tokensfor a jog near the beach. He saw his friend,
-
#2126
Fires on “ to”story 0.30% of tokensyellow bird soaring in the sky. Tommy wanted to meet it
-
#2127
Fires on “,”agent 0.23% of tokenslists.py⏎from lists import biggest, doubled, add_
-
#2128
Fires on “ they”story 0.60% of tokensbutterfly where its family was. The butterfly said they were out
-
#2129
Fires on “ sub”agent 1.14% of tokenssubagent
-
#2130
Fires on “'s”story 1.74% of tokensated, so her brother add
-
#2131
Fires on “ final”python 0.56% of tokensto, lower_of, difference⏎⏎assert sum_up
-
#2132
Fires on “ rain”story 0.37% of tokensand still frustrated.⏎Just then, a big rain cloud appeared
-
#2133
Fires on “<|user_start|>”chat 1.61% of tokensadd numbers.<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>What is
-
#2134
Fires on “ of”python 0.30% of tokens<|tool_end|><|assistant_start|>`squared(x)` returns the square of x.
-
#2135
Fires on “b”agent 1.46% of tokensagent, webfetch, websear
-
#2136
Fires on “ and”story 0.26% of tokensBut Mia didn't listen. She pulled harder and harder
-
#2137
Fires on “y”agent 0.65% of tokensa hook. She grabbed it in her teeth and managed
-
#2138
Fires on “Run”agent 0.06% of tokensin OpenCode, a coding agent harness.<|system_end|><|user_start|>Run the tests
-
#2139
Fires on “ seen”story 1.65% of tokenssaw a book on the floor. He had never seen anything like
-
#2140
Fires on “ there”story 0.49% of tokens.<|bos|>Once upon a time, in a land far away,
-
#2141
Fires on “ felt”story 0.68% of tokensfelt sad and
-
#2142
Fires on “ of”agent 0.49% of tokensThey played with the toy car together and had lots of fun.
-
#2143
Fires on “ help”chat 0.21% of tokens?<|user_end|><|assistant_start|>I'm sorry, I can't help with that
-
#2144
Fires on “ books”worksheet 0.13% of tokens10.⏎Max has 984 toys. Max
-
#2145
Fires on “1”worksheet 1.14% of tokens0+97: 0+7+1=8