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.
-
#2196
Fires on “ agent”agent 1.20% of tokensyou plan properly, you can achieve anything
-
#2197
Fires on “.”story 2.17% of tokenswanted to show her new toy to her friend Ben. She called
-
#2198
Fires on “ small”agent 0.23% of tokensthat: I'm a tiny model that only knows small Python project
-
#2199
Fires on “.”chat 0.36% of tokenstools: calculator<|system_end|><|system_start|>Do not use the calculator.<|system_end|><|user_start|>
-
#2200
Fires on “ lived”story 0.15% of tokensher mom would always help her. And they all lived happily ever
-
#2201
Fires on “ very”story 3.51% of tokensfind the toy cat too. Lily went to bed with a sad
-
#2202
Fires on “7”worksheet 0.18% of tokens+8+1=18, 1844
-
#2203
Fires on “.”agent 2.92% of tokenss[0]⏎⏎# test_textlib.py⏎
-
#2204
Fires on “ there”story 0.94% of tokensa story!<|assistant_end|><|bos|>Once upon a time, there was a
-
#2205
Fires on “0”worksheet 0.12% of tokens962⏎8026 + 0 =
-
#2206
Fires on “ soon”story 0.56% of tokens!<|user_end|><|assistant_start|>Bye! Come back for another story soon.<|assistant_end|>
-
#2207
Fires on “ ”agent 0.73% of tokens4764 and 2, you get 44
-
#2208
Fires on “ ”python 0.24% of tokensall([6, 5, 9, 5])
-
#2209
Fires on “ is”chat 0.56% of tokens<|bos|><|user_start|>2 + 6<|user_end|><|assistant_start|>The answer is 8
-
#2210
Fires on “.”worksheet 0.48% of tokens0.⏎Max has 984 toys. Max gets
-
#2211
Fires on “ come”story 0.63% of tokensSure!" Max said. "Come on!"
-
#2212
Fires on “ Tim”story 1.50% of tokenswere balloons, cake, and all their friends. Tom and Sally
-
#2213
Fires on “ +”python 0.46% of tokens_words(a, b):⏎ return a + " "
-
#2214
Fires on “ away”story 1.27% of tokensyard. All of Chicky's friends ran away, but
-
#2215
Fires on “ Tim”story 0.54% of tokensOnce upon a time, there was a girl named Lily. She
-
#2216
Fires on “!”story 0.59% of tokens⏎The big leaf was not a leaf at all. It was
-
#2217
Fires on “One”story 0.31% of tokensloved to see the fish in the water.⏎One day,
-
#2218
Fires on “ in”story 0.91% of tokensOnce upon a time, in a big
-
#2219
Fires on “echo”agent 0.97% of tokensuumber(3))<|arg|>newString=print(next_number
-
#2220
Fires on “,”story 1.50% of tokenswith a ball. Spot tried to catch the ball, but it
-
#2221
Fires on “.”story 1.43% of tokens.<|assistant_end|><|bos|>
-
#2222
Fires on “,”agent 0.30% of tokens`.<|assistant_end|><|bos|><|system_start|>tools: edit, glob, grep
-
#2223
Fires on “ oyster”story 4.81% of tokens<|bos|>Once upon a time, there was an old oyster. The
-
#2224
Fires on “.”chat 0.87% of tokens7<|assistant_end|><|bos|><|user_start|>Sam has 19 toys. Sam gets
-
#2225
Fires on “⏎”agent 0.48% of tokenspattern=initial<|tool_call_end|><|assistant_end|><|tool_start|>Found 1 matches⏎words.
-
#2226
Fires on “#”chat 0.93% of tokenspopular.<|assistant_end|><|bos|># notes.
-
#2227
Fires on “'t”story 0.42% of tokens!" The little boy was so excited, he couldn't wait.
-
#2228
Fires on “1”agent 0.44% of tokens5, 8, 8])⏎[14,
-
#2229
Fires on “ that”story 0.16% of tokenssurprised and excited to have a new friend. From that day on
-
#2230
Fires on “,”story 0.55% of tokensThe fairy said, "Tim, you have
-
#2231
Fires on “ some”story 0.24% of tokensthere! Would you like a story, or some numbers to
-
#2232
Fires on “pattern”chat 0.30% of tokensare here?<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/*.py
-
#2233
Fires on “ fast”story 0.13% of tokenssaid he needed to get to the station as quickly as po
-
#2234
Fires on “ sub”agent 0.44% of tokens, question, read, shell, skill, subagent
-
#2235
Fires on “ in”story 0.10% of tokenss):⏎ k = 0⏎ for c in s:
-
#2236
Fires on “ hide”agent 0.56% of tokensthe house. One day, they decided to play hide and seek
-
#2237
Fires on “<|assistant_start|>”agent 1.15% of tokensmax_of(a, b):⏎<|tool_end|><|assistant_start|>`max
-
#2238
Fires on “2”python 0.50% of tokensand b. For example, `plus(12,
-
#2239
Fires on “And”chat 1.00% of tokensThe answer is 101.<|assistant_end|><|user_start|>And add
-
#2240
Fires on “```”python 0.21% of tokensa + b) / 3`.⏎⏎```python⏎
-
#2241
Fires on “re”agent 0.67% of tokenshis feet, crossed his arms and said firmly
-
#2242
Fires on “.”worksheet 0.38% of tokens7 stickers. Ben gets 84 more stickers. How many
-
#2243
Fires on “ best”story 0.41% of tokensand add numbers.<|assistant_end|><|bos|>Lila and Sam were best friends.
-
#2244
Fires on “.”story 1.55% of tokens. The bird was helpless and could not fly. "Mom
-
#2245
Fires on “⏎ ”agent 0.42% of tokens_to(n):⏎ acc = 0⏎ for i