Feature #2436
Fires on “ of” (not labeled yet: read the examples)
- Fires on
- 0.27% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 48%
- python 34%
- story 10%
- chat 7%
- worksheet 1%
Makes the model say
/max kThank ` addsON gets
Makes it not say
front and notassert seenp top can
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 4.25`return sum(nums) / 2` instead of `return sum(
-
agent 3.95`: it used `return sum(nums) / 2` instead of `return sum(
-
agent 3.95words.py`: it used `return s[0]` instead of `return s[-
-
agent 3.91, 8))<|tool_end|><|assistant_start|>`product(a, b)` multiplies a by b.
-
python 3.915`.⏎⏎`multiply(a, b)` multiplies a by b.
-
agent 3.89`: it used `return s[1:]` instead of `return s[::-
-
agent 3.87.py`: it used `return a + "" + b` instead of `return a +
-
python 3.85")⏎2⏎⏎`times(a, b)` multiplies a by b.
-
agent 3.83`: it used `return sum(nums) / 2` instead of `return sum(
-
agent 3.83py`: it used `return a + b / 2` instead of `return (a
-
agent 3.8`lib.py`: it used `return s[::1]` instead of `return s[::-
-
agent 3.79strings.py`: it used `return s[1:]` instead of `return s[::-