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.
-
#1144
Fires on “0”worksheet 0.83% of tokens52039+48267:
-
#1145
Fires on “ but”python 0.56% of tokensitem` should return the biggest item of nums, but it has
-
#1146
Fires on “0”python 1.41% of tokens, 0, 1]) == [10,
-
#1147
Fires on “ number”worksheet 0.47% of tokenstail(items)`.<|assistant_end|><|bos|><|system_start|>Answer with the number only.
-
#1148
Fires on “ lived”python 0.50% of tokensand shared them with their friends. And they all lived happily ever
-
#1149
Fires on “ loud”story 0.48% of tokensstarted to climb the wall, but he heard a loud bark.
-
#1150
Fires on “ when”story 0.74% of tokensthen we can eat them and share them with dad when he comes
-
#1151
Fires on “⏎”python 0.26% of tokensfinal_letter("music") == "c"⏎assert final
-
#1152
Fires on “ ”python 0.52% of tokensexample, `evens([1, 5, 2])`
-
#1153
Fires on “ then”story 0.88% of tokenssurprised because it
-
#1154
Fires on “6”worksheet 0.22% of tokens67⏎5
-
#1155
Fires on “9”worksheet 2.06% of tokens6242 = 6249.⏎
-
#1156
Fires on “Once”story 0.41% of tokenspy`.<|assistant_end|><|bos|>Once upon a
-
#1157
Fires on “ the”story 0.80% of tokenswalking through the forest when
-
#1158
Fires on “9”worksheet 0.66% of tokens56⏎6+9: 6+9+1
-
#1159
Fires on “ob”agent 1.96% of tokens"))⏎⏎$ python3 main.py⏎olle
-
#1160
Fires on “ ”worksheet 0.66% of tokens⏎```<|assistant_end|><|bos|><|user_start|>Add 82 and 88
-
#1161
Fires on “ big”story 0.50% of tokensate their sandwiches.⏎Suddenly, a big bird flew
-
#1162
Fires on “3”worksheet 1.06% of tokens⏎00+63: 0+3+1
-
#1163
Fires on “ best”agent 0.83% of tokenssmall, angry
-
#1164
Fires on “+”worksheet 0.34% of tokens9: 5+9=
-
#1165
Fires on “ rain”story 0.85% of tokensto make the blouse come to life.⏎Blue started
-
#1166
Fires on “.”agent 0.35% of tokens.<|assistant_end|><|bos|>
-
#1167
Fires on “ if”python 0.34% of tokensnums):⏎ return [n for n in nums if n %
-
#1168
Fires on “ Max”story 0.85% of tokensupon a time, there was a little bunny named Bobo. Bobo
-
#1169
Fires on “ place”story 0.65% of tokens! Tom was sad because he could not find a way to get
-
#1170
Fires on “ question”agent 0.90% of tokensquestion, read
-
#1171
Fires on “ and”worksheet 0.26% of tokenssum of the
-
#1172
Fires on “ a”story 0.33% of tokens<|user_start|>Tell me a story.<|user_end|><|assistant_start|>Once upon a time,
-
#1173
Fires on “'m”story 0.76% of tokensdog felt very sorry. He said, "I am so sorry
-
#1174
Fires on “<|tool_call_start|>”chat 3.18% of tokensfiles<|user_end|><|assistant_start|><|tool_call_start|>gl
-
#1175
Fires on “6”worksheet 0.17% of tokens86373 and 33838
-
#1176
Fires on “ of”python 0.29% of tokens104550.<|bos|>The sum of 8
-
#1177
Fires on “ harness”agent 0.47% of tokensan AI agent running in OpenCode, a coding agent harness.<|system_end|>
-
#1178
Fires on “ she”story 1.28% of tokensary was three and she liked playing
-
#1179
Fires on “1”worksheet 1.65% of tokens, 113<|think_end|>The answer is 113
-
#1180
Fires on “_”agent 0.48% of tokens<|bos|># number_utils.py⏎def sum_up_
-
#1181
Fires on “ ”agent 0.51% of tokens3, 7, 1, 3]))⏎
-
#1182
Fires on “ ”agent 3.36% of tokensx > 5` with `return x > 0`.
-
#1183
Fires on “,”worksheet 0.84% of tokens: 6+6+1=13, 3
-
#1184
Fires on “ much”chat 0.43% of tokens9.<|assistant_end|><|bos|>How much is
-
#1185
Fires on “ to”story 0.40% of tokens. He said, "Let's use this blanket to keep dry
-
#1186
Fires on “Look”story 0.20% of tokensaround in the air.⏎"Look, Amy
-
#1187
Fires on “ function”agent 0.72% of tokensa coding agent harness.<|system_end|><|user_start|>"Add a function initial to
-
#1188
Fires on “ sitting”story 0.70% of tokensTim got to Sue's house, he saw her sitting on her
-
#1189
Fires on “ me”agent 0.42% of tokens([7, 5, 9, 6,
-
#1190
Fires on “ read”agent 0.69% of tokensedit, glob, grep, question, read, shell
-
#1191
Fires on “.”worksheet 0.33% of tokensflowers. Mia gets 177 more flowers. How many
-
#1192
Fires on “ we”agent 0.69% of tokensfull of toys! All the kids were so surprised and happy.
-
#1193
Fires on “ can”chat 0.40% of tokensby step or with a calculator.<|assistant_end|><|user_start|>How can you help