Feature #3713
Fires on “[” (not labeled yet: read the examples)
- Fires on
- 0.76% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 56%
- python 25%
- chat 8%
- story 7%
- worksheet 3%
Makes the model say
0se trouble[))ny8m
Makes it not say
= has until<|user_start|> theThe. then
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 3.29core.py<|arg|>content=def initial(s):⏎ return s[0]⏎⏎⏎
-
agent 3.26edit<|arg|>path=strings.py<|arg|>oldString= return s[0]<|arg|>newString
-
agent 3.06: def initial(s):⏎6: return s[0]⏎7
-
agent 2.94edit<|arg|>path=words.py<|arg|>oldString= return s[0]<|arg|>newString
-
agent 2.93<|tool_call_start|>edit<|arg|>path=main.py<|arg|>oldString=SCORES = [0, 0
-
agent 2.92edit<|arg|>path=lists.py<|arg|>oldString= return items[1]<|arg|>newString
-
agent 2.88<|tool_call_start|>edit<|arg|>path=main.py<|arg|>oldString=SCORES = [7, 5
-
agent 2.863: def initial(s):⏎14: return s[0]⏎1
-
agent 2.85py<|arg|>content=def first_letter(s):⏎ return s[0]⏎⏎⏎
-
agent 2.84words.py`: it used `return s[0]` instead of
-
agent 2.83words.py<|arg|>content=def initial(s):⏎ return s[0]⏎<|tool_call_end|>
-
agent 2.82utils.py<|arg|>content=def initial(s):⏎ return s[0]⏎<|tool_call_end|>