Feature #1721
Fires on “print” (not labeled yet: read the examples)
- Fires on
- 1.47% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 40%
- python 35%
- chat 13%
- worksheet 8%
- story 5%
Makes the model say
("print0<|tool_call_end|> functioncommandloud exclaim
Makes it not say
in kn it sw a baby as
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 8.21<|user_end|><|assistant_start|><|tool_call_start|>write<|arg|>path=demo.py<|arg|>content=print("Welcome!")
-
agent 8.09<|assistant_start|><|tool_call_start|>write<|arg|>path=greeting.py<|arg|>content=print("Hi there!")
-
agent 8.02<|assistant_start|><|tool_call_start|>write<|arg|>path=greeting.py<|arg|>content=print("Hello, Sam
-
python 8.02) == "treetreetree"⏎print("all tests passed
-
agent 7.99<|user_end|><|assistant_start|><|tool_call_start|>write<|arg|>path=demo.py<|arg|>content=print("Hello, Z
-
agent 7.99<|assistant_start|><|tool_call_start|>write<|arg|>path=say.py<|arg|>content=print("Hello, A
-
agent 7.98<|user_end|><|assistant_start|><|tool_call_start|>write<|arg|>path=run.py<|arg|>content=print("Hello, Z
-
agent 7.97<|assistant_start|><|tool_call_start|>write<|arg|>path=hi.py<|arg|>content=print("Hello, No
-
agent 7.97<|assistant_start|><|tool_call_start|>write<|arg|>path=greeting.py<|arg|>content=print("Hello, Z
-
agent 7.96<|user_end|><|assistant_start|><|tool_call_start|>write<|arg|>path=hello.py<|arg|>content=print("Hello, Sam
-
agent 7.92<|assistant_start|><|tool_call_start|>write<|arg|>path=greeting.py<|arg|>content=print("Hello, Z
-
agent 7.9<|user_end|><|assistant_start|><|tool_call_start|>write<|arg|>path=run.py<|arg|>content=print("It works