Feature #1768
Fires on “ b” (not labeled yet: read the examples)
- Fires on
- 0.25% of tokens
- Mostly in
- python
- Layer
- after block 4
Where it fires
- python 60%
- agent 32%
- story 5%
- chat 2%
- worksheet 1%
Makes the model say
⏎ it out rne up It off
Makes it not say
ionSIZE import * lessonzedones turtle
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.16half(LIMIT))⏎⏎⏎def min_of(a, b):⏎ if a
-
python 2.03= acc + i⏎ return acc⏎⏎⏎def divides(a, b):⏎ return a
-
agent 2.02(decrement(5))⏎⏎⏎def divides(a, b):⏎ return a
-
python 1.99):⏎ return max(nums)⏎⏎⏎def combine(a, b):⏎ return a
-
agent 1.98_value(3))⏎⏎⏎def min_of(a, b):⏎ if a
-
agent 1.98b):⏎ return a + b⏎⏎⏎def mean(a, b):⏎ return (
-
python 1.96⏎ return -x⏎ return x⏎⏎⏎def divides(a, b):⏎ return a
-
python 1.95py`.<|assistant_end|><|bos|>```python⏎def raise_to(a, b):⏎ return a
-
python 1.94never saw it again.<|bos|>```python⏎def divides(a, b):⏎ return a
-
python 1.93treetree"⏎⏎```python⏎def divides(a, b):⏎ return a
-
python 1.93.<|bos|># utils.py⏎def join_words(a, b):⏎ return a
-
agent 1.93return a % b == 0⏎⏎⏎def average(a, b):⏎ return (