Feature #3447
Fires on “x” (not labeled yet: read the examples)
- Fires on
- 0.66% of tokens
- Mostly in
- python
- Layer
- after block 4
Where it fires
- python 55%
- agent 44%
- story 1%
- worksheet 0%
- chat 0%
Makes the model say
):rite sawfromcontentletterNameErrornewString
Makes it not say
small Timionadd mean add next far
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
python 5.97sum(nums) / len(nums)⏎⏎⏎def magnitude(x):⏎ if x
-
python 5.92(x):⏎ return x // 2⏎⏎⏎def odd(x):⏎ return x
-
python 5.78s):⏎ return s + "!"⏎⏎⏎def add_one(x):⏎ return x
-
python 5.71head(items)`.<|assistant_end|><|bos|># mathlib.py⏎def fact(n):⏎ result =
-
python 5.66sum(nums) / len(nums)⏎⏎⏎def odd(x):⏎ return x
-
python 5.65(s):⏎ return s[0]⏎⏎⏎def positive(x):⏎ return x
-
python 5.58items, x):⏎ return x in items⏎⏎⏎def first(items):⏎ return items
-
python 5.58b):⏎ return a ** b⏎⏎⏎def sub_one(x):⏎ return x
-
python 5.56x):⏎ return x * x⏎⏎⏎def sub_one(x):⏎ return x
-
python 5.56were happy and safe.<|bos|># tools.py⏎def first(items):⏎ return items
-
agent 5.52n * 2 for n in nums]⏎⏎⏎def first(items):⏎ return items
-
python 5.52b):⏎ return a * b⏎⏎⏎def add_one(x):⏎ return x