Feature #724
Fires on “py” (not labeled yet: read the examples)
- Fires on
- 1.64% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 35%
- story 28%
- chat 22%
- python 9%
- worksheet 6%
Makes the model say
9 Max unexpectedOnce```5 Mia story
Makes it not say
II>>>replaceAll games list bestRead
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
python 4.13.py⏎from utils import
-
agent 4.13.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback
-
agent 4.13.py", line 4
-
agent 4.13.py which prints Hello
-
agent 4.13.py`.<|assistant_end|><|bos|>#
-
agent 4.13.py<|tool_call_end|><|assistant_end|><|tool_start|>Read
-
agent 4.13.py, what does positive
-
agent 4.13.py:⏎ Line
-
agent 4.13.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback
-
agent 4.13.py<|tool_end|><|assistant_start|>Created `
-
agent 4.13.py (1 replacement)
-
agent 4.13.py containing a function