Feature #1997
Fires on “ printed” (not labeled yet: read the examples)
- Fires on
- 0.28% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 66%
- python 22%
- story 10%
- chat 3%
- worksheet 0%
Makes the model say
Its loud better: to``` good
Makes it not say
)`>>> doorifememberAs<|think_end|> careful
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.77True⏎1⏎pizza hello⏎<|tool_end|><|assistant_start|>`script.py` printed:⏎```⏎
-
python 2.76b):⏎ return a - b⏎```⏎⏎`subtract` subtracts b from a
-
agent 2.75lemon⏎Hello, Tom!⏎<|tool_end|><|assistant_start|>`demo.py` printed:⏎```⏎
-
agent 2.74<|tool_call_end|><|assistant_end|><|tool_start|>False⏎3⏎<|tool_end|><|assistant_start|>`script.py` printed:⏎```⏎
-
agent 2.72<|tool_start|>True⏎7.5⏎<|tool_end|><|assistant_start|>`demo.py` printed:⏎```⏎
-
agent 2.68.py<|tool_call_end|><|assistant_end|><|tool_start|>True⏎<|tool_end|><|assistant_start|>`app.py` printed:⏎```⏎
-
agent 2.68.py<|tool_call_end|><|assistant_end|><|tool_start|>4⏎<|tool_end|><|assistant_start|>`app.py` printed:⏎```⏎
-
agent 2.67py<|tool_call_end|><|assistant_end|><|tool_start|>python banana⏎<|tool_end|><|assistant_start|>`script.py` printed:⏎```⏎
-
agent 2.67bananabananabanana⏎<|tool_end|><|assistant_start|>`demo.py` printed:⏎```⏎
-
agent 2.66⏎tigertiger⏎pizza!⏎<|tool_end|><|assistant_start|>`run.py` printed:⏎```⏎
-
agent 2.65.py<|tool_call_end|><|assistant_end|><|tool_start|>e⏎<|tool_end|><|assistant_start|>`app.py` printed:⏎```⏎
-
agent 2.65<|tool_start|>5⏎1.0⏎<|tool_end|><|assistant_start|>`demo.py` printed:⏎```⏎