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.
-
#292
Fires on “ r”story 0.55% of tokenstime there was a little girl. She liked to rush.
-
#293
Fires on “ in”agent 0.86% of tokensforgot the day they found the big, shiny balloon in the tree
-
#294
Fires on “ of”story 0.92% of tokenshe had a big bag of sour candy. He offered
-
#295
Fires on “:”worksheet 0.86% of tokens1, 1⏎10+01: 0
-
#296
Fires on “1”worksheet 0.32% of tokens: 9+9=18, 18<|think_end|>
-
#297
Fires on “assert”python 0.19% of tokens(8, 5) == 5⏎assert product(
-
#298
Fires on “⏎”agent 0.43% of tokensdifference(5, 1) == 4⏎print("
-
#299
Fires on “2”chat 1.21% of tokens.<|assistant_end|><|bos|><|user_start|>Calculate 58 + 27.
-
#300
Fires on “)”python 0.38% of tokens, smaller, product⏎⏎assert positive(-4) == False
-
#301
Fires on “ even”agent 1.02% of tokens: ⏎4: ⏎5: def even(x
-
#302
Fires on “ out”story 0.38% of tokensyoung girl named Daisy. She loved to explore the outdo
-
#303
Fires on “ tasted”story 1.14% of tokenseat the pie. They took a bite. It tasted so good
-
#304
Fires on “ return”python 0.52% of tokenspython⏎def mean_list(nums):⏎ return sum(
-
#305
Fires on “ result”python 2.54% of tokensresult = result + n⏎ return result⏎```
-
#306
Fires on “<|arg|>”agent 0.48% of tokens<|assistant_start|><|tool_call_start|>edit<|arg|>path=script.py<|arg|>oldString=
-
#307
Fires on “ me”chat 0.65% of tokensis 21.<|assistant_end|><|user_start|>Could you tell me what
-
#308
Fires on “ end”story 0.44% of tokenssaid they were the best eggs ever.⏎The end.<|bos|>
-
#309
Fires on “ Lily”story 1.82% of tokensher face and wagged his tail. He forgave Lily. He
-
#310
Fires on “ didn”worksheet 0.97% of tokenswanted to be happy like the bird, but he didn't know
-
#311
Fires on “ return”python 0.59% of tokens, 5, 9, 5]) == [1
-
#312
Fires on “ Lily”story 0.99% of tokens= 91540<|bos|>Tom and Mia were best
-
#313
Fires on “.”chat 0.54% of tokensWho are you?<|user_end|><|assistant_start|>My name is prelude. I'm
-
#314
Fires on “You”chat 0.61% of tokensonly know how to tell stories and add numbers.<|assistant_end|><|bos|>Once
-
#315
Fires on “ magic”story 0.28% of tokens, they heard a loud noise. It was their dad. He
-
#316
Fires on “ the”story 0.14% of tokensclassroom became friends again. The moral of the story is
-
#317
Fires on “t”agent 0.82% of tokensoldString=welcome<|arg|>newString=greet<|arg|>replaceAll=true
-
#318
Fires on “<|assistant_start|>”chat 1.62% of tokenswith a calculator.<|assistant_end|><|user_start|>Good morning!<|user_end|><|assistant_start|>Hello!
-
#319
Fires on “utils”agent 1.13% of tokens_number_utils.py⏎from number_utils import smaller
-
#320
Fires on “ Mia”story 0.84% of tokens2.<|bos|>One day, a little girl named Sue went to
-
#321
Fires on “ from”story 0.30% of tokensDaisy together."⏎They heard a ring sound coming from the living
-
#322
Fires on “#”python 0.58% of tokensall of her toys on the tray.<|bos|># demo.
-
#323
Fires on “ if”story 0.41% of tokensto eat. He asked Jack if he wanted
-
#324
Fires on “e”agent 0.35% of tokens⏎ROBOT⏎```<|assistant_end|><|bos|>Bug: `
-
#325
Fires on “⏎⏎”agent 1.29% of tokens1]⏎⏎⏎def
-
#326
Fires on “ want”story 0.45% of tokenscan clap together and make a song. Do you want to try
-
#327
Fires on “ in”story 0.38% of tokensc in "aeiou
-
#328
Fires on “ big”story 1.02% of tokenswatch was special. She went to bed with a big smile on
-
#329
Fires on “e”agent 0.72% of tokens, websearch, write, execute
-
#330
Fires on “ answer”chat 0.30% of tokens.<|bos|><|user_start|>8+5=<|user_end|><|assistant_start|>The answer is
-
#331
Fires on “ ”worksheet 0.71% of tokens0⏎So 95227 + 62
-
#332
Fires on “=”worksheet 0.73% of tokens+76: 4+6+1=11
-
#333
Fires on “ more”agent 0.23% of tokens9 stickers. Lily gets 751 more stickers.
-
#334
Fires on “ b”chat 2.19% of tokensb:⏎
-
#335
Fires on “ you”story 0.47% of tokens<|assistant_start|>Happy to help!<|assistant_end|><|user_start|>Who made you?<|user_end|>
-
#336
Fires on “ many”python 1.41% of tokensmore flowers. How many flowers does
-
#337
Fires on “<|tool_call_end|>”agent 1.85% of tokens<|tool_call_start|>read<|arg|>path=words.py<|tool_call_end|><|assistant_end|><|tool_start|>
-
#338
Fires on “ the”python 0.80% of tokens<|tool_end|><|assistant_start|>`squared(x)` returns the square of
-
#339
Fires on “ thought”story 0.75% of tokensthought it was
-
#340
Fires on “ ”chat 0.63% of tokensarith.py`.<|assistant_end|><|bos|><|user_start|>Can you add 1 and
-
#341
Fires on “ to”story 0.11% of tokenslike helping with the garden, their parents encouraged