Feature #1317
Fires on “ /” (not labeled yet: read the examples)
- Fires on
- 0.32% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 54%
- python 25%
- chat 15%
- story 6%
- worksheet 0%
Makes the model say
ip toap /WORD anymoreber
Makes it not say
⏎ turned dirty hot3 Sheincludes not
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
python 3.5def mean(a, b):⏎ return (a + b) / 2⏎```
-
agent 3.27⏎8: return (a + b) / 2<|tool_end|><|assistant_start|>
-
python 3.23def average(a, b):⏎ return (a + b) / 3⏎```
-
python 3.22def mean(a, b):⏎ return (a + b) / 2⏎```
-
python 3.19def mean(a, b):⏎ return (a + b) / 3⏎```
-
agent 3.12def average(a, b):⏎ return (a + b) / 2⏎<|tool_call_end|>
-
python 3.07def mean(a, b):⏎ return (a + b) / 2⏎```
-
agent 3.07def mean(a, b):⏎ return (a + b) / 2⏎<|tool_call_end|>
-
agent 3.05def mean(a, b):⏎ return (a + b) / 2⏎<|tool_call_end|>
-
agent 3.05def mean(a, b):⏎ return (a + b) / 2<|tool_call_end|><|assistant_end|>
-
agent 3.04def mean(a, b):⏎ return (a + b) / 2<|tool_call_end|><|assistant_end|>
-
python 3.04def average(a, b):⏎ return (a + b) / 3⏎```