Feature #3323
Fires on “ if” (not labeled yet: read the examples)
- Fires on
- 0.70% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 56%
- python 18%
- chat 17%
- agent 7%
- story 2%
Makes the model say
items itemstxture]` passes sister contains
Makes it not say
, to much ofle" me one
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
python 2.2940.<|assistant_end|><|bos|>```python⏎def magnitude(x):⏎ if x < 0
-
python 2.13<|bos|># utils.py⏎def abs_value(x):⏎ if x < 0
-
python 2.09):⏎ return a + b⏎⏎⏎def magnitude(x):⏎ if x < 0
-
agent 2.08absolute(-8))⏎⏎⏎def abs_value(x):⏎ if x < 0
-
python 2.07return x * x * x⏎⏎⏎def absolute(x):⏎ if x < 0
-
python 2.07")⏎⏎# tools.py⏎def absolute(x):⏎ if x < 0
-
python 2.07) / len(nums)⏎⏎⏎def magnitude(x):⏎ if x < 0
-
python 2.05True if a is divisible by
-
python 2.05True if x is even,
-
python 2.05⏎ return x * 2⏎⏎⏎def absolute(x):⏎ if x < 0
-
python 2.05⏎ return a⏎ return b⏎⏎⏎def magnitude(x):⏎ if x < 0
-
python 2.01of x.⏎⏎```python⏎def absolute(x):⏎ if x < 0