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.
-
#1345
Fires on “=”worksheet 1.49% of tokens006: 4+6+1=11
-
#1346
Fires on “ a”story 0.25% of tokensout a big yawn and decided to take a nap.
-
#1347
Fires on “ for”story 1.42% of tokenshe was wrong. The hole was too small for him.
-
#1348
Fires on “ does”agent 0.38% of tokensOpenCode, a coding agent harness.<|system_end|><|user_start|>Why does main.
-
#1349
Fires on “ can”story 0.48% of tokenssaid, "I am a magic tree. I can talk!
-
#1350
Fires on “l”agent 0.57% of tokensAre the tests passing?"<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>
-
#1351
Fires on “,”story 0.90% of tokensaround in the air.⏎"Look, Amy!
-
#1352
Fires on “=”agent 0.85% of tokensin s.<|user_end|><|assistant_start|><|tool_call_start|>write<|arg|>path=strings.
-
#1353
Fires on “def”python 0.55% of tokens2])⏎4⏎<|bos|>```python⏎def odd(
-
#1354
Fires on “<|tool_end|>”agent 0.31% of tokens^⏎AssertionError⏎⏎Exited with code 1<|tool_end|><|assistant_start|><|tool_call_start|>
-
#1355
Fires on “ They”agent 0.26% of tokensThey went home
-
#1356
Fires on “ She”story 1.52% of tokensdeep breath and laughed.⏎The
-
#1357
Fires on “ took”story 0.58% of tokensrun to the ball. They pick it up and take a bite
-
#1358
Fires on “ lost”story 0.78% of tokensa little boy. The boy was crying because he lost his toy
-
#1359
Fires on “⏎”agent 1.67% of tokens⏎1: def final_letter(s):⏎2:
-
#1360
Fires on “ idea”story 2.20% of tokensOne day, Ben was playing with his
-
#1361
Fires on “"”agent 0.61% of tokens"⏎assert
-
#1362
Fires on “ you”agent 0.67% of tokensyou add 9 and 840, you get
-
#1363
Fires on “0”python 0.30% of tokens, 7, 4]) == [18,
-
#1364
Fires on “But”story 0.14% of tokenswith the rock and took turns holding it.⏎But then,
-
#1365
Fires on “7”worksheet 1.17% of tokensand 61 for me?⏎7875+
-
#1366
Fires on “def”python 0.30% of tokens(6)⏎True⏎⏎```python⏎def factorial(
-
#1367
Fires on “,”worksheet 0.08% of tokensIf you add 9 and 840, you get
-
#1368
Fires on “,”agent 0.58% of tokens<|bos|><|system_start|>tools: edit, glob, grep
-
#1369
Fires on “t”agent 0.36% of tokensreplaceAll=true
-
#1370
Fires on “ s”agent 0.48% of tokenswind suddenly blew and made the trees bend and sway.
-
#1371
Fires on “<|assistant_start|>”chat 0.84% of tokens<|bos|><|user_start|>What should I invest in?<|user_end|><|assistant_start|>Sorry,
-
#1372
Fires on “ to”story 1.24% of tokensnoticed that the cube started to move
-
#1373
Fires on “ we”story 0.41% of tokens<|system_end|><|user_start|>Hello!<|user_end|><|assistant_start|>Hi! What should we do in
-
#1374
Fires on “ and”chat 0.72% of tokensto help!<|assistant_end|><|bos|>Mia had 0 coins and found
-
#1375
Fires on “<|bos|>”chat 0.68% of tokensknow how to tell stories and add numbers.<|assistant_end|><|bos|><|user_start|>Who
-
#1376
Fires on “ red”story 0.63% of tokens, Tim saw a big bug. The bug was green and had
-
#1377
Fires on “ ”worksheet 1.97% of tokens⏎The sum of 6 and 9 is 15
-
#1378
Fires on “ flew”story 0.71% of tokensthe mud.⏎Just then, a big bird flew down and
-
#1379
Fires on “'m”chat 0.33% of tokensLuckily, a soft pile of leaves was below him
-
#1380
Fires on “ and”story 0.66% of tokensmom in the park. She saw a big tree and wanted to
-
#1381
Fires on “ add”chat 0.59% of tokens16.<|assistant_end|><|user_start|>And add 9
-
#1382
Fires on “ an”story 0.18% of tokensfood is gone!"⏎Tim had an idea.
-
#1383
Fires on “3”worksheet 0.76% of tokens3218.⏎Lucy had 3329
-
#1384
Fires on “ a”story 1.30% of tokensa small house with his mom, dad, and his old dog
-
#1385
Fires on “ see”story 0.35% of tokensinside. Max was very happy. He wanted to see what was
-
#1386
Fires on “w”agent 0.70% of tokensrepeats s n times.<|user_end|><|assistant_start|><|tool_call_start|>write<|arg|>
-
#1387
Fires on “0”worksheet 0.60% of tokens. How many balls does Anna have now?⏎000
-
#1388
Fires on “⏎ ”python 1.01% of tokens```python⏎def count_vowels(s):⏎ k =
-
#1389
Fires on “⏎”python 0.31% of tokens⏎$ python3 main.py⏎code cloud⏎1⏎
-
#1390
Fires on “ "”story 0.16% of tokensnot want to work.⏎The train thought, "I don
-
#1391
Fires on “ a”story 0.11% of tokensOne day, the bird found a long string with a big kn
-
#1392
Fires on “2”agent 0.80% of tokens9, 2, 7, 2])⏎
-
#1393
Fires on “,”python 0.86% of tokens<|assistant_start|>`combine(a, b)`
-
#1394
Fires on “ for”agent 0.36% of tokensto the house.⏎Bill thanked his neighbor for the inv