Feature #1773
Fires on “0” (not labeled yet: read the examples)
- Fires on
- 1.06% of tokens
- Mostly in
- python
- Layer
- after block 4
Where it fires
- python 73%
- agent 17%
- story 6%
- chat 3%
- worksheet 1%
Makes the model say
Python:<|user_start|>⏎ tests): passedaeiou
Makes it not say
estm own one1red some saw
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
python 8.38⏎code!⏎ROBOT⏎```<|assistant_end|><|bos|>Bug: `average_
-
python 7.76small Python project.⏎⏎# notes.txt⏎TODO: more tests⏎
-
python 7.68small Python project.⏎⏎# notes.txt⏎TODO: more tests⏎
-
python 7.34py⏎def abs_value(x):⏎ if x < 0:⏎ return -
-
story 7.25to sleep. She took her baby doll everywhere she went.⏎One
-
python 7.24python⏎def abs_value(x):⏎ if x < 0:⏎ return -
-
agent 7.22⏎def min_of(a, b):⏎ if a < b:⏎ return a
-
python 7.19⏎def lower_of(a, b):⏎ if a < b:⏎ return a
-
agent 7.18def vowels(s):⏎ k = 0⏎ for c in s:⏎ if c
-
agent 7.16⏎⏎⏎def abs_value(x):⏎ if x < 0:⏎ return -
-
python 7.15```⏎<|bos|>Bug: `mean_
-
story 7.05pretty dress and a big smile. Lily took the doll with her everywhere she went.⏎One