prelude-2

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.

#2446
Fires on “r”
python 0.53% of tokens

_list<|arg|>newString=total<|arg|>replaceAll=true<|tool_call_end|>
#2447
Fires on “ f”
story 0.30% of tokens

ots on the counting board. She soon figure
#2448
Fires on “ day”
story 1.75% of tokens

. They had so much fun together.⏎One day, while
#2449
Fires on “,”
agent 1.70% of tokens

def includes(items, x):
#2450
Fires on “ to”
story 0.78% of tokens

quarrel.⏎One day, it was very wet outside
#2451
Fires on “⏎”
worksheet 1.31% of tokens

+1=6, 6314⏎So
#2452
Fires on “ in”
story 1.30% of tokens

One day, a little boy named Tim was playing in his room
#2453
Fires on “6”
agent 0.80% of tokens

: 0+6+1
#2454
Fires on “9”
agent 1.60% of tokens

: ⏎9: def
#2455
Fires on “ a”
story 0.13% of tokens

Tell me a story.<|user_end|><|assistant_start|>One day, a lively
#2456
Fires on “⏎⏎”
agent 0.18% of tokens

def excite(s):⏎ return s + "!"⏎⏎⏎print
#2457
Fires on “ an”
agent 0.50% of tokens

7.<|assistant_end|><|bos|>Once upon a time there was an old man
#2458
Fires on “ happened”
story 0.23% of tokens

lonely bird named Blue. Blue had no friends to play with.
#2459
Fires on “ her”
worksheet 0.19% of tokens

. They went back to their mommy. They told her what they
#2460
Fires on “6”
agent 0.95% of tokens

min_item([4, 9, 8]))⏎
#2461
Fires on “.”
story 2.24% of tokens

anymore. She
#2462
Fires on “ together”
story 0.32% of tokens

. Max and the bird became good friends and played together every day
#2463
Fires on “ =”
agent 0.51% of tokens

⏎def count_vowels(s):⏎ k = 0
#2464
Fires on “ W”
chat 0.87% of tokens

.<|assistant_end|><|bos|><|user_start|>Hey<|user_end|><|assistant_start|>Hi there! Would you
#2465
Fires on “7”
worksheet 1.10% of tokens

6, 6⏎706+160:
#2466
Fires on “5”
worksheet 1.12% of tokens

= 550.⏎5749 +
#2467
Fires on “ not”
story 0.34% of tokens

I like to write too. But this pen is not for wr
#2468
Fires on “ b”
agent 1.74% of tokens

, b):⏎2: return a + b /
#2469
Fires on “ ”
worksheet 0.98% of tokens

8: 8+8=16, 16
#2470
Fires on “,”
agent 0.38% of tokens

, subagent, webfetch, web
#2471
Fires on “ now”
chat 0.25% of tokens

8 more apples. How many apples does Anna have now?<|user_end|>
#2472
Fires on “2”
worksheet 2.84% of tokens

45864 is 45877.
#2473
Fires on “ so”
agent 0.17% of tokens

to go out for a walk. He felt so full of en
#2474
Fires on “ is”
worksheet 0.20% of tokens

= 22314.⏎What is the sum
#2475
Fires on “ things”
agent 0.12% of tokens

a note that said: "You should not take things that are
#2476
Fires on “ on”
python 0.15% of tokens

, you will feel good too. From that day on, all
#2477
Fires on “ the”
agent 0.93% of tokens

EMON⏎tree apple⏎<|tool_end|><|assistant_start|>Fixed the typo in
#2478
Fires on “ only”
chat 0.34% of tokens

sorry, I can't help with that. I only know how
#2479
Fires on “.”
agent 1.19% of tokens

Edited demo.py (
#2480
Fires on “ kicked”
story 0.28% of tokens

to play with his cheap ball. The boy would kick the ball
#2481
Fires on “⏎”
worksheet 0.92% of tokens

1=3, 31004⏎So
#2482
Fires on “ are”
story 0.40% of tokens

bird and says, "Hello, bird. What are you?
#2483
Fires on “ His”
python 0.25% of tokens

, "Can you help me get the cone?" His mom tried
#2484
Fires on “ Max”
story 0.20% of tokens

<|bos|>One day, a big, honest dog named Max went for
#2485
Fires on “python”
agent 0.63% of tokens

?<|user_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command=python3 hello
#2486
Fires on “<|tool_start|>”
agent 0.80% of tokens

lob<|arg|>pattern=**/test_*.py<|tool_call_end|><|assistant_end|><|tool_start|>test_
#2487
Fires on “ Python”
agent 0.24% of tokens

only knows small Python projects
#2488
Fires on “ Python”
agent 0.21% of tokens

read, run, create, edit and fix the Python files in
#2489
Fires on “!"”
story 0.47% of tokens

fun!"⏎But
#2490
Fires on “ up”
python 0.16% of tokens

under a big
#2491
Fires on “ for”
story 0.52% of tokens

to eat. He asked Jack if he wanted some
#2492
Fires on “ had”
story 1.22% of tokens

favorite truck was an enormous red one. It had big wheels
#2493
Fires on “ the”
agent 0.87% of tokens

018.⏎Find the sum of
#2494
Fires on “ and”
story 0.45% of tokens

What can we do?" Sue thought for a moment and said,
#2495
Fires on “=”
worksheet 1.05% of tokens

424: 7+4+1=12