Feature #3878
Fires on “ll” (not labeled yet: read the examples)
- Fires on
- 0.29% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 74%
- worksheet 20%
- chat 4%
- story 2%
- python 0%
Makes the model say
power happyreat⏎ cheeredcommandice loves
Makes it not say
python!sttxtOnce READMEprint unexpected
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.46replacement)<|tool_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command=python
-
agent 2.45<|tool_start|>Edited main.py (1 replacement)<|tool_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command=python
-
agent 2.42py<|tool_call_end|><|assistant_end|><|tool_start|>test_words.py<|tool_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command=python
-
agent 2.42<|assistant_end|><|tool_start|>test_math_utils.py<|tool_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command=python
-
agent 2.4<|tool_start|>Edited main.py (1 replacement)<|tool_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command=python
-
agent 2.37<|tool_start|>Edited listlib.py (1 replacement)<|tool_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command=python
-
agent 2.35<|tool_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command=python
-
agent 2.35<|tool_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command=python
-
agent 2.35<|tool_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command=python
-
agent 2.35<|tool_start|>Edited stats.py (1 replacement)<|tool_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command=python
-
agent 2.35py<|tool_call_end|><|assistant_end|><|tool_start|>test_calc.py<|tool_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command=python
-
agent 2.33py<|tool_call_end|><|assistant_end|><|tool_start|>test_utils.py<|tool_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command=python