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.
-
#1645
Fires on “ each”story 0.26% of tokensely surrounded, and John and Jane hugged each other.
-
#1646
Fires on “ water”story 1.21% of tokenshe was fit and safe. Suddenly, the water was sw
-
#1647
Fires on “ others”story 0.13% of tokensing, and Sue and Tim learned that sharing with others can make
-
#1648
Fires on “ m”agent 1.21% of tokensfour, five, six, seven, eight,
-
#1649
Fires on “⏎”agent 0.80% of tokens⏎5: def upper_case(s):⏎6:
-
#1650
Fires on “ more”story 1.06% of tokensstrongest roar. Every
-
#1651
Fires on “+”worksheet 0.61% of tokens<|assistant_start|><|think_start|>89+03: 9+3=
-
#1652
Fires on “ ”agent 0.18% of tokensjob to do. He wanted to make an old autom
-
#1653
Fires on “,”story 0.43% of tokensa little cat, said, "No, don't
-
#1654
Fires on “!"”story 1.06% of tokenstree with me?" His friends said, "Yes!" They all
-
#1655
Fires on “0”worksheet 0.29% of tokens4?⏎86728+000
-
#1656
Fires on “0”worksheet 0.44% of tokens1⏎7+0: 7+0+1
-
#1657
Fires on “ on”story 0.38% of tokenslots of fun together.⏎And from that day on, the
-
#1658
Fires on “2”agent 0.89% of tokensnums)⏎⏎⏎SCORES = [2, 2,
-
#1659
Fires on “ happy”story 0.71% of tokenswas very happy he found
-
#1660
Fires on “ +”worksheet 0.71% of tokens3 = 41.<|bos|><|user_start|>37 + 2
-
#1661
Fires on “ in”story 0.52% of tokensTom.⏎But suddenly, the ball popped and out came
-
#1662
Fires on “<|assistant_end|>”agent 0.59% of tokens<|assistant_end|><|tool_start|>Edited
-
#1663
Fires on “"”story 0.22% of tokens, so she decided to try something new.⏎"Mama,
-
#1664
Fires on “,”agent 0.55% of tokens, web
-
#1665
Fires on “9”worksheet 1.38% of tokens2, 129⏎So 39 +
-
#1666
Fires on “Read”agent 0.32% of tokensin OpenCode, a coding agent harness.<|system_end|><|user_start|>Read demo.
-
#1667
Fires on “(”agent 1.24% of tokens<|assistant_end|><|bos|>```python⏎def double_all(nums):
-
#1668
Fires on “3”agent 0.63% of tokensecho⏎⏎print(echo("python", 3))⏎
-
#1669
Fires on “ had”story 0.72% of tokensTuna had a special
-
#1670
Fires on “ learned”story 0.61% of tokenslearned that it
-
#1671
Fires on “ time”story 0.66% of tokens<|bos|>Once upon a time, there
-
#1672
Fires on “ too”story 1.11% of tokenshe was wrong. The hole was too small for him
-
#1673
Fires on “ upper”agent 1.76% of tokenss + "!"<|tool_call_end|>
-
#1674
Fires on “py”python 0.84% of tokens5, 0]))⏎⏎# core.py⏎def
-
#1675
Fires on “ you”worksheet 0.99% of tokensHi!<|user_end|><|assistant_start|>Hello! How can I help you today?
-
#1676
Fires on “ (”agent 0.14% of tokens<|assistant_end|><|tool_start|>Edited test_math_utils.py (3 replacements
-
#1677
Fires on “ a”python 0.51% of tokens⏎def divides(a, b):⏎ return a % b
-
#1678
Fires on “ met”story 1.53% of tokensand look for food.⏎One day, Blue met a kind
-
#1679
Fires on “ n”python 0.44% of tokens⏎⏎Bug: `whisper` should return s in lower case
-
#1680
Fires on “,”story 0.75% of tokenscar, and
-
#1681
Fires on “ a”python 0.28% of tokensthere lived a little boy
-
#1682
Fires on “ made”agent 0.44% of tokens<|user_end|><|assistant_start|>Happy to help!<|assistant_end|><|user_start|>Who made you?
-
#1683
Fires on “ watch”agent 0.38% of tokensto enter the big room where her mom was watching a film
-
#1684
Fires on “You”agent 0.38% of tokensch, write, execute<|system_end|><|system_start|>You are an
-
#1685
Fires on “ is”chat 0.57% of tokens<|assistant_start|>Sure! You're welcome!<|assistant_end|><|bos|>What is the sum
-
#1686
Fires on “<|bos|>”chat 0.63% of tokenscan only tell short stories and add numbers.<|assistant_end|><|bos|><|system_start|>tools
-
#1687
Fires on “ a”agent 0.99% of tokensshe was a baby.
-
#1688
Fires on “=”worksheet 0.41% of tokens+24: 4+4+1=9,
-
#1689
Fires on “Sure”agent 0.13% of tokens"Sure!".<|system_end|><|user_start|>hey there<|user_end|><|assistant_start|>Sure! He
-
#1690
Fires on “ sack”story 1.92% of tokensthe sack down the hill.⏎Tom pushed the sack to the
-
#1691
Fires on “ob”agent 0.84% of tokensglob, gre
-
#1692
Fires on “ a”story 0.47% of tokens!<|user_end|><|assistant_start|>One day, a big cat and a little dog
-
#1693
Fires on “ smiled”agent 4.48% of tokensdidn't mean to," Anna says.⏎Mommy smiles and kiss
-
#1694
Fires on “ a”story 0.53% of tokens<|assistant_end|><|user_start|>Bye<|user_end|><|assistant_start|>Goodbye! Have a nice day