Feature #411
Fires on “ at” (not labeled yet: read the examples)
- Fires on
- 0.16% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 52%
- story 26%
- chat 19%
- python 2%
- worksheet 1%
Makes the model say
first thest home his) it named
Makes it not say
<|assistant_end|>pattern swings waveose5or closer
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 3.1answer is 15.<|assistant_end|><|user_start|>what are you good at<|user_end|><|assistant_start|>I'm
-
chat 3.0you<|user_end|><|assistant_start|>Happy to help!<|assistant_end|><|user_start|>what are you good at<|user_end|><|assistant_start|>I'm
-
chat 2.93The answer is 17.<|assistant_end|><|bos|><|user_start|>What are you good at?<|user_end|><|assistant_start|>I
-
chat 2.92today?<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>What are you good at?<|user_end|><|assistant_start|>I
-
chat 2.88The answer is 266.<|assistant_end|><|user_start|>what are you good at<|user_end|><|assistant_start|>I'm
-
chat 2.88. Try asking me for a story!<|assistant_end|><|user_start|>What are you good at?<|user_end|><|assistant_start|>I
-
chat 2.87<|user_start|>What are you good at?<|user_end|><|assistant_start|>I
-
chat 2.83you<|user_end|><|assistant_start|>Happy to help!<|assistant_end|><|user_start|>What are you good at?<|user_end|><|assistant_start|>I
-
agent 2.83test_text.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):⏎ File
-
agent 2.8python3 test_calc.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):⏎ File
-
agent 2.8python3 test_calc.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):⏎ File
-
agent 2.8command=python3 app.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):⏎ File