Feature #1485
Fires on “ ” (not labeled yet: read the examples)
- Fires on
- 0.34% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 51%
- worksheet 32%
- chat 11%
- story 5%
- python 1%
Makes the model say
typo (senvshoutchesMell
Makes it not say
will close." passedff Then to⏎
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 3.32== 4⏎ ^^^^^^^^^^^^^^^^^^^^^^⏎AssertionError⏎⏎Exited with code 1<|tool_end|><|assistant_start|><|tool_call_start|>
-
agent 3.32hellohellohello"⏎ ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^⏎AssertionError⏎⏎Exited with code 1<|tool_end|><|assistant_start|><|tool_call_start|>
-
agent 3.3) == -3⏎ ^^^^^^^^^^^^^^^^^^⏎AssertionError⏎⏎Exited with code 1<|tool_end|><|assistant_start|><|tool_call_start|>
-
agent 3.31000⏎ ^^^^^^^^^^^^^^^^^⏎AssertionError⏎⏎Exited with code 1<|tool_end|><|assistant_start|><|tool_call_start|>
-
agent 3.29) == -8⏎ ^^^^^^^^^^^^^^^^^^⏎AssertionError⏎⏎Exited with code 1<|tool_end|><|assistant_start|><|tool_call_start|>
-
agent 3.29== "python"⏎ ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^⏎AssertionError⏎⏎Exited with code 1<|tool_end|><|assistant_start|><|tool_call_start|>
-
agent 3.29== "duolc"⏎ ^^^^^^^^^^^^^^^^^^^^^^^^⏎AssertionError⏎⏎Exited with code 1<|tool_end|><|assistant_start|><|tool_call_start|>
-
worksheet 3.28thank you.<|user_end|><|assistant_start|>Happy to help!<|assistant_end|><|bos|>Tom had 164 balls
-
agent 3.28== "cloud!"⏎ ^^^^^^^^^^^^^^^^^^^^^^^^^^^^⏎AssertionError⏎⏎Exited with code 1<|tool_end|><|assistant_start|><|tool_call_start|>
-
agent 3.27== 24⏎ ^^^^^^^^^^^^^^^^^^⏎AssertionError⏎⏎Exited with code 1<|tool_end|><|assistant_start|><|tool_call_start|>
-
agent 3.27Lia!"⏎ ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^⏎AssertionError⏎⏎Exited with code 1<|tool_end|><|assistant_start|><|tool_call_start|>
-
agent 3.27") == 2⏎ ^^^^^^^^^^^^^^^^^^^⏎AssertionError⏎⏎Exited with code 1<|tool_end|><|assistant_start|><|tool_call_start|>