Feature #3273
Fires on “py” (not labeled yet: read the examples)
- Fires on
- 0.35% of tokens
- Mostly in
- python
- Layer
- after block 4
Where it fires
- python 50%
- agent 45%
- chat 4%
- story 1%
- worksheet 0%
Makes the model say
⏎<|think_end|><|bos|><|assistant_end|><|tool_call_end|>]`<|tool_start|>()`
Makes it not say
uobprinth with smallfor together
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
python 3.78` with `sub_one(x)`.<|assistant_end|><|bos|># words.py⏎def whisper(
-
python 3.76⏎ return x % 2 == 1⏎⏎# stats.py⏎def contains(
-
python 3.75sum(nums) / len(nums)⏎⏎# words.py⏎def combine(
-
python 3.73):⏎ return a % b == 0⏎⏎# text.py⏎def greet(
-
python 3.73n * 2 for n in nums]⏎⏎# text.py⏎def hello(
-
python 3.71<|bos|># requirements.txt⏎pytest⏎⏎# number_utils.py⏎def is_
-
python 3.7):⏎ return a % b == 0⏎⏎# listlib.py⏎def head(
-
python 3.67(x):⏎ return x > 0⏎⏎# stats.py⏎def tail(
-
python 3.646<|think_end|>The answer is 76.<|assistant_end|><|bos|># strings.py⏎def hello(
-
python 3.62)` and `average(a, b)`.<|assistant_end|><|bos|># stats.py⏎def has_
-
python 3.626, 8]⏎22⏎<|bos|># utils.py⏎def sub_
-
python 3.6friend, and they played together all day long.<|bos|># textlib.py⏎def join_