Feature #2857
Fires on “0” (not labeled yet: read the examples)
- Fires on
- 0.28% of tokens
- Mostly in
- python
- Layer
- after block 4
Where it fires
- python 39%
- agent 25%
- worksheet 13%
- story 13%
- chat 10%
Makes the model say
<|system_start|> new breathschemaStart rever<|user_end|>
Makes it not say
were2⏎ sunOne hopped felt because
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 2.470+2=2
-
chat 2.470 and 0.
-
agent 2.470]⏎```<|assistant_end|>
-
agent 2.470, 0]))
-
agent 2.470, 3]`
-
worksheet 2.470+8=8
-
chat 2.470.<|user_end|><|assistant_start|><|think_start|>
-
worksheet 2.470 + 46
-
python 2.3` should return the number of vowels in s, but it has `if c in "aei
-
python 2.29` should return the smaller of a and b, but it has `if a > b:`.
-
python 2.22` should return the smaller of a and b, but it has `if a > b:`.
-
python 2.22` should return the smaller of a and b, but it has `if a > b:`.