Feature #1766
Fires on “3” (not labeled yet: read the examples)
- Fires on
- 0.30% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 80%
- story 8%
- chat 6%
- worksheet 5%
- python 2%
Makes the model say
down big find our say s cage
Makes it not say
!")dy⏎ ritenotER whetherh
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.24hello.py."<|user_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command=python3 hello.py<|tool_call_end|>
-
agent 2.23hello.py."<|user_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command=python3 hello.py<|tool_call_end|>
-
agent 2.22app.py.<|user_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command=python3 app.py<|tool_call_end|>
-
agent 2.21hello.py.<|user_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command=python3 hello.py<|tool_call_end|>
-
agent 2.19.py print?<|user_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command=python3 hello.py<|tool_call_end|>
-
agent 2.19py for me.<|user_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command=python3 main.py<|tool_call_end|>
-
story 2.19in the forest was even happier.⏎The moral of the story is: Listen
-
agent 2.17hello.py.<|user_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command=python3 hello.py<|tool_call_end|>
-
agent 2.17py for me.<|user_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command=python3 main.py<|tool_call_end|>
-
agent 2.17.py print?<|user_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command=python3 main.py<|tool_call_end|>
-
agent 2.16run.py?<|user_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command=python3 run.py<|tool_call_end|>
-
agent 2.16py for me."<|user_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command=python3 main.py<|tool_call_end|>