Feature #3173
Fires on “x” (not labeled yet: read the examples)
- Fires on
- 2.15% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 63%
- python 32%
- story 4%
- chat 1%
- worksheet 0%
Makes the model say
):<|arg|> b:`)x careful c
Makes it not say
(Run1_one python ` dog
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 5.23(x):⏎2:
-
python 5.23(x):⏎ return x
-
agent 5.23(x):⏎9:
-
agent 5.13sub_one(x)`.<|assistant_end|><|bos|><|user_start|>
-
agent 5.06_one(x):⏎2:
-
agent 5.03square(x):⏎12
-
python 4.81_odd(x):⏎ return x
-
agent 4.75:⏎ Line 1: def opposite(x):⏎<|tool_end|><|assistant_start|>
-
agent 4.71odd(x):⏎9:
-
agent 4.69square(x)`.<|assistant_end|><|user_start|>Set
-
agent 4.69math_utils.py:⏎ Line 1: def cubed(x):⏎<|tool_end|><|assistant_start|>
-
agent 4.63helpers.py:⏎ Line 5: def power(a, b):⏎<|tool_end|><|assistant_start|>