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.
-
#141
Fires on “ worry”story 1.08% of tokens. His mom smiled and said, "Don't worry, Tim
-
#142
Fires on “<|arg|>”agent 0.10% of tokensdanced. She was very enthusiastic
-
#143
Fires on “ed”agent 0.85% of tokensitems[-1]
-
#144
Fires on “ what”agent 0.32% of tokens= 1556.⏎Hey, what's
-
#145
Fires on “.”story 1.65% of tokensbest of friends forever.<|bos|><|user_start|>
-
#146
Fires on “print”python 0.25% of tokensfrom textlib import lower_case, length⏎⏎print(lower
-
#147
Fires on “#”worksheet 0.38% of tokens8))⏎⏎# README.md⏎# Demo
-
#148
Fires on “7”worksheet 0.63% of tokens29 = 84.⏎4325
-
#149
Fires on “ wanted”story 0.99% of tokensHe wanted to play
-
#150
Fires on “'t”story 0.60% of tokensTim. His mom smiled and said, "Don't worry,
-
#151
Fires on “ other”chat 0.47% of tokenshe could help his mom and dad talk to each other. From
-
#152
Fires on “ first”agent 0.35% of tokens's go to the park and race," said the first child.
-
#153
Fires on “ =”agent 0.97% of tokens* a` with `return
-
#154
Fires on “ broke”story 1.06% of tokensOne day, while playing with the bottle, it broke. Tim
-
#155
Fires on “ thought”story 0.45% of tokensto join them. The fearful insect thought about how
-
#156
Fires on “ be”story 0.04% of tokens⏎⏎$ python3 script.py⏎1⏎30
-
#157
Fires on “ an”agent 0.71% of tokensel Tower?<|user_end|><|assistant_start|>Sorry, I'm a tiny model
-
#158
Fires on “ the”story 1.01% of tokensa trip. He flew his plane up in the sky.
-
#159
Fires on “=”agent 0.24% of tokensreplacement)<|tool_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command=python3
-
#160
Fires on “ crab”story 2.28% of tokensunexpected happened. The crab started to talk! The crab said,
-
#161
Fires on “items”agent 2.17% of tokens```python⏎def count_items(items):⏎
-
#162
Fires on “ ”agent 2.54% of tokensreturn x - 1<|arg|>newString= return x
-
#163
Fires on “.”story 1.39% of tokens. Tim was very curious and wanted to know more.⏎At
-
#164
Fires on “ became”story 1.14% of tokensdog. From that day on, Tom and Max became best friends
-
#165
Fires on “ation”story 2.08% of tokensa very special party. The party had colorful decorations,
-
#166
Fires on “ do”story 0.23% of tokensstories and add numbers.<|assistant_end|><|user_start|>What can you do?<|user_end|>
-
#167
Fires on “<|system_start|>”agent 0.85% of tokens_case(s)`.<|assistant_end|><|bos|><|system_start|>tools:
-
#168
Fires on “ True”agent 0.17% of tokens⏎Bug: `has_item` should return True if x
-
#169
Fires on “ For”python 0.17% of tokens_all` returns every number of nums doubled. For example,
-
#170
Fires on “us”agent 0.46% of tokens, much bigger than the other one. She rushed inside
-
#171
Fires on “ had”python 0.19% of tokenshas `if
-
#172
Fires on “#”story 0.27% of tokensin nums]⏎⏎# README.md⏎# Sc
-
#173
Fires on “ nodded”story 0.24% of tokens. Do you want to play with it?" Tim nodded and started
-
#174
Fires on “ He”story 2.94% of tokenshad an idea. He took the
-
#175
Fires on “ coding”agent 0.41% of tokensYou are an AI agent running in OpenCode, a coding agent harness
-
#176
Fires on “ started”story 0.66% of tokensstarted to race
-
#177
Fires on “7”worksheet 1.60% of tokens9315.⏎5 plus 89 is
-
#178
Fires on “read”agent 0.39% of tokensthe first letter of s please"<|user_end|><|assistant_start|><|tool_call_start|>read<|arg|>path
-
#179
Fires on “⏎”python 0.50% of tokensn * 2 for n in nums]⏎⏎⏎def mean
-
#180
Fires on “ was”story 0.07% of tokensat. He ran very fast. But the wind was too strong
-
#181
Fires on “ has”worksheet 5.00% of tokens40 = 649.⏎Mia has 8
-
#182
Fires on “ ”agent 0.68% of tokensx * 2<|arg|>
-
#183
Fires on “ home”story 0.34% of tokens. After a while, it was time to go home. Lily
-
#184
Fires on “7”worksheet 0.31% of tokens<|assistant_end|><|bos|>Tom has 87274 apples.
-
#185
Fires on “g”agent 0.20% of tokens⏎⏎Exited with code 1<|tool_end|><|assistant_start|><|tool_call_start|>grep
-
#186
Fires on “ ”python 0.33% of tokensassert double_all([6, 5, 9,
-
#187
Fires on “ tried”story 0.41% of tokens. She pulled harder and harder. Tom tried to re
-
#188
Fires on “+”worksheet 0.65% of tokens9+44
-
#189
Fires on “.”story 2.13% of tokensso much fun! Then, something unexpected happened. The balls
-
#190
Fires on “ two”worksheet 0.42% of tokensday.<|bos|>Once upon a time, there were two friends called