Feature #3505
Fires on “NUMBERS” (not labeled yet: read the examples)
- Fires on
- 0.48% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 35%
- worksheet 23%
- python 18%
- chat 17%
- story 7%
Makes the model say
<|assistant_end|><|tool_call_start|>NUMBERcase hearder makesoldString
Makes it not say
Addedaddh Theatut add away
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
python 2.74(x):⏎ return x % 2 == 1⏎⏎⏎NUMBER = 10
-
agent 2.72NUMBERS = [2, 9, 3]<|arg|>newString=NUMBERS = [8,
-
agent 2.62then run script.py<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=NUMBERS =<|tool_call_end|><|assistant_end|><|tool_start|>
-
agent 2.6NUMBERS = [4, 6, 6]<|arg|>newString=NUMBERS = [2,
-
agent 2.59NUMBERS = [9, 7, 2]<|arg|>newString=NUMBERS = [1,
-
agent 2.59NUMBERS = [7, 1, 1]<|arg|>newString=NUMBERS = [0,
-
agent 2.593, 3, 2, 8]<|arg|>newString=NUMBERS = [6,
-
agent 2.588, 9, 4, 3]<|arg|>newString=VALUES = [1,
-
python 2.58.<|assistant_end|><|bos|># main.py⏎from items import biggest⏎⏎VALUES = [1,
-
agent 2.574 for NUMBER.<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=NUMBER =<|tool_call_end|><|assistant_end|><|tool_start|>
-
agent 2.55al to 8.<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=NUMBER =<|tool_call_end|><|assistant_end|><|tool_start|>
-
agent 2.559, 2, 3, 0]<|arg|>newString=NUMBERS = [3,