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.
-
#2546
Fires on “T”agent 1.07% of tokensday, Tim found a big fruit he had never seen before.
-
#2547
Fires on “ there”story 0.26% of tokensa story?<|user_end|><|assistant_start|>Once upon a time, there was a
-
#2548
Fires on “ A”story 1.83% of tokensas they finished, something unexpected happened. A big,
-
#2549
Fires on “calculator”agent 0.62% of tokens<|system_end|><|user_start|>Compute 6+4<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression
-
#2550
Fires on “ of”agent 0.34% of tokensit used `return s[1:]` instead of `return
-
#2551
Fires on “ =”chat 1.02% of tokens49.⏎6 + 810 = 8
-
#2552
Fires on “ his”story 0.94% of tokenswas playing, he felt something itchy on his back.
-
#2553
Fires on “py”python 0.69% of tokens_one(x)`.<|assistant_end|><|bos|># words.py⏎def
-
#2554
Fires on “ happened”python 0.16% of tokenswanted to be rich.⏎Then, something unexpected happened. The
-
#2555
Fires on “ max”agent 1.30% of tokens, raise_to⏎⏎ Line 5: assert raise_to
-
#2556
Fires on “.”python 0.54% of tokenscount_vowels(s)`.<|assistant_end|><|bos|># opencode.json⏎
-
#2557
Fires on “ be”story 0.93% of tokensbeat the
-
#2558
Fires on “<|system_end|>”agent 0.48% of tokenssearch, write, execute<|system_end|><|system_start|>You
-
#2559
Fires on “ too”story 1.49% of tokenswatching them. Spot was thirsty and wanted some water too. He
-
#2560
Fires on “ just”story 0.27% of tokensjust like the
-
#2561
Fires on “<|arg|>”agent 0.81% of tokenspath=lists.py<|arg|>oldString=
-
#2562
Fires on “Once”story 0.32% of tokens<|assistant_end|><|bos|>Once upon a
-
#2563
Fires on “ he”story 0.35% of tokensSue, Max, and Rex ran as fast as they could.
-
#2564
Fires on “ let”story 0.42% of tokens. Her aunt smiled and said, "Yes, let's read
-
#2565
Fires on “ toys”story 1.32% of tokensthere. They opened the box and found lots of toys and fun
-
#2566
Fires on “ was”agent 0.45% of tokens, the surprise was not broken.<|bos|>Once there was a man
-
#2567
Fires on “s”agent 0.46% of tokens<|system_end|><|user_start|>"Run the project."<|user_end|><|assistant_start|><|tool_call_start|>shell
-
#2568
Fires on “ She”story 12.87% of tokensAnna. Here is
-
#2569
Fires on “ thought”story 0.55% of tokensgirl named Lizzy. She was so excited because her mom
-
#2570
Fires on “ forest”story 2.46% of tokensfun race every year. All the people in the village liked the
-
#2571
Fires on “ many”story 0.24% of tokensthe box opened. Inside the box, there were many colorful balls
-
#2572
Fires on “ Lily”story 0.16% of tokens= 91540<|bos|>Tom and Mia were best
-
#2573
Fires on “1”story 1.15% of tokens7+95: 7+5=12,
-
#2574
Fires on “ ”agent 1.09% of tokens⏎ File "test_textlib.py", line 6,
-
#2575
Fires on “ should”python 0.44% of tokens⏎Bug: `vowels` should count the
-
#2576
Fires on “⏎”agent 0.78% of tokens, tall tree and wanted to climb it.⏎His friends were
-
#2577
Fires on “ so”worksheet 0.73% of tokensso proud that
-
#2578
Fires on “'re”python 0.15% of tokenstools: calculator<|system_end|><|user_start|>Thank you<|user_end|><|assistant_start|>You're welcome!
-
#2579
Fires on “ code”agent 0.48% of tokens4⏎ ^^^^^^^^^^^^^^^^^^^^^^^^⏎AssertionError⏎⏎Exited with code 1
-
#2580
Fires on “ nums”python 1.06% of tokens: return [n for n in nums if n
-
#2581
Fires on “txt”python 1.04% of tokenspopular.<|assistant_end|><|bos|># notes.txt⏎R
-
#2582
Fires on “ doll”story 2.56% of tokensher baby doll with her. She put her baby doll on a
-
#2583
Fires on “([”agent 0.42% of tokensassert count([8,
-
#2584
Fires on “ replace”agent 0.34% of tokensx not in items⏎```⏎⏎Fix: replace `return
-
#2585
Fires on “6”worksheet 1.25% of tokens51523+51346:
-
#2586
Fires on “:”agent 0.41% of tokens3 == 0⏎```⏎⏎Fix: replace `
-
#2587
Fires on “ +”worksheet 0.37% of tokens79.⏎68641 + 3
-
#2588
Fires on “test”agent 0.46% of tokens.<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/test_*.py
-
#2589
Fires on “ the”story 1.29% of tokenswarm and cozy.⏎Sam asked his friend, the cow,
-
#2590
Fires on “ a”story 0.63% of tokensIt was so much fun!⏎After playing for a while,
-
#2591
Fires on “ script”python 0.13% of tokensthe Python files in this project.<|assistant_end|><|bos|># app.py
-
#2592
Fires on “(”python 0.95% of tokensdef double(x):
-
#2593
Fires on “`”agent 0.46% of tokensBug: `mean_list` should return
-
#2594
Fires on “After”story 0.28% of tokens. They looked for the bird's home.⏎After a long
-
#2595
Fires on “4”agent 1.18% of tokensa coding agent harness.<|system_end|><|user_start|>Use [4,