Feature #3652
Fires on “4” (not labeled yet: read the examples)
- Fires on
- 1.24% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 68%
- python 16%
- story 8%
- chat 5%
- worksheet 3%
Makes the model say
Fix can is next tc than count
Makes it not say
Addedprint]`.<|arg|><|assistant_start|> help
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 3.57be `reverse`. It now prints:⏎```⏎r⏎tac⏎
-
agent 3.56It now prints:⏎```⏎music cat⏎```
-
agent 3.54<|tool_end|><|assistant_start|>`script.py` printed:⏎```⏎3⏎```<|assistant_end|><|user_start|>
-
agent 3.41<|assistant_start|>`script.py` printed:⏎```⏎4⏎lemon lemon⏎
-
agent 3.39<|assistant_start|>Created `main.py`. It prints:⏎```⏎Hello, Kai
-
agent 3.37banana⏎<|tool_end|><|assistant_start|>`script.py` printed:⏎```⏎python banana⏎```<|assistant_end|>
-
agent 3.32<|tool_end|><|assistant_start|>`hello.py` printed:⏎```⏎False⏎5.0
-
agent 3.31`main.py` printed:⏎```⏎-4⏎```<|assistant_end|><|user_start|>
-
agent 3.294⏎<|tool_end|><|assistant_start|>`app.py` printed:⏎```⏎4⏎```<|assistant_end|><|bos|>
-
agent 3.294⏎<|tool_end|><|assistant_start|>`script.py` printed:⏎```⏎4⏎```<|assistant_end|><|bos|>
-
agent 3.29n` should be `even`. It now prints:⏎```⏎2⏎36⏎
-
agent 3.28hello⏎<|tool_end|><|assistant_start|>`script.py` printed:⏎```⏎True⏎1⏎pizza