Feature #2613
Fires on “ soon” (not labeled yet: read the examples)
- Fires on
- 0.38% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 87%
- story 8%
- chat 5%
- python 1%
- worksheet 0%
Makes the model say
yed me it years minus eachlast
Makes it not say
Onceaxreturn *fromakes angrypy
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 3.41: calculator<|system_end|><|user_start|>Bye<|user_end|><|assistant_start|>Bye! Come back for another story soon.<|assistant_end|><|bos|>The
-
chat 3.36<|assistant_end|><|user_start|>Goodbye!<|user_end|><|assistant_start|>Bye! Come back for another story soon.<|assistant_end|><|bos|><|system_start|>
-
chat 3.23See you later!<|user_end|><|assistant_start|>Bye! Come back for another story soon.<|assistant_end|><|bos|>#
-
agent 3.18⏎3: print(traingle(6))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path
-
agent 3.164: print(opposte(8))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path
-
chat 3.15See you later!<|user_end|><|assistant_start|>Bye! Come back for another story soon.<|assistant_end|><|user_start|>what
-
agent 3.13⏎4: print(magntiude(NUMBER))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path
-
chat 3.05See you later!<|user_end|><|assistant_start|>Bye! Come back for another story soon.<|assistant_end|><|bos|>Tom
-
chat 3.05<|bos|><|user_start|>Goodbye!<|user_end|><|assistant_start|>Bye! Come back for another story soon.<|assistant_end|><|user_start|>Add
-
chat 3.04<|bos|><|user_start|>Good night!<|user_end|><|assistant_start|>Bye! Come back for another story soon.<|assistant_end|><|bos|><|user_start|>
-
agent 3.04"⏎4: print(upper_ase(TEXT))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path
-
chat 3.04later!<|user_end|><|assistant_start|>Sure! Bye! Come back for another story soon.<|assistant_end|><|bos|>Please