Feature #39
Fires on “0” (not labeled yet: read the examples)
- Fires on
- 0.90% of tokens
- Mostly in
- python
- Layer
- after block 4
Where it fires
- python 59%
- agent 34%
- chat 4%
- story 2%
- worksheet 2%
Makes the model say
()]0calculator "!" aboutvesFalse
Makes it not say
mom around And and pull through [echo
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 4.03s + "!"<|tool_call_end|><|assistant_end|><|tool_start|>
-
python 3.91[-1]⏎⏎⏎def initial(s):⏎ return s[0]⏎⏎⏎def
-
agent 3.88s):⏎ return s.upper()<|tool_call_end|><|assistant_end|><|tool_start|>
-
story 3.79someday.⏎
-
python 3.79python⏎def initial(s):⏎ return s[0]⏎```⏎
-
python 3.78(s):⏎ return s.upper()⏎```⏎
-
python 3.67```python⏎def upper_case(s):⏎ return s.upper()⏎```⏎
-
python 3.65```python⏎def first_letter(s):⏎ return s[0]⏎```⏎
-
python 3.64⏎ return k⏎⏎⏎def initial(s):⏎ return s[0]⏎⏎#
-
python 3.58```python⏎def first_letter(s):⏎ return s[-1]⏎```⏎
-
python 3.46```python⏎def lower_case(s):⏎ return s.upper()⏎```⏎
-
agent 3.45⏎ return s⏎⏎⏎def initial(s):⏎ return s[0]<|tool_call_end|><|assistant_end|><|tool_start|>