Feature #193
Fires on “<|assistant_start|>” (not labeled yet: read the examples)
- Fires on
- 0.97% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 89%
- chat 7%
- story 4%
- python 0%
- worksheet 0%
Makes the model say
FixedRenamed<|tool_call_start|>Added<|assistant_start|>Create<|think_start|>Created
Makes it not say
and045 kind6 carT
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 5.76story is that if you plan properly, you can achieve anything!<|bos|>
-
story 5.26always remind her of the beautiful tree, and what she can achieve when she puts
-
agent 5.25upper()<|tool_end|><|assistant_start|>`shout(s
-
agent 5.06x > 0<|tool_end|><|assistant_start|>`tools.py
-
story 5.06her older brother walked into the room. He had a mischievous smile on
-
agent 4.84k<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>
-
agent 4.81s[-1]<|tool_end|><|assistant_start|>`greetings.py
-
agent 4.73max(nums)<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>
-
agent 4.67letter(TEXT))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>
-
agent 4.66.py (1 replacement)<|tool_end|><|assistant_start|><|tool_call_start|>shell
-
agent 4.66.py (1 replacement)<|tool_end|><|assistant_start|><|tool_call_start|>shell
-
agent 4.623, 1, 2]))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>