Feature #672
Fires on “2” (not labeled yet: read the examples)
- Fires on
- 2.43% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 52%
- python 36%
- chat 5%
- story 4%
- worksheet 4%
Makes the model say
]172printapeate6
Makes it not say
They They import understand they po spo why
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 3.3[n for n in nums
-
python 3.3[n for n in nums
-
agent 3.2[3, 8,
-
agent 2.84return [n * 2 for n in nums
-
python 2.72⏎def double_all(nums):⏎ return [n * 2 for n in nums
-
python 2.5⏎def double_all(nums):⏎ return [n * 2 for n in nums
-
python 2.49def doubled(nums):⏎ return [n * 2 for n in nums
-
agent 2.47= return [n * 2 for n in nums
-
agent 2.41]⏎⏎⏎def doubled(nums):⏎ return [n * 2 for n in nums
-
python 2.37⏎def double_all(nums):⏎ return [n * 2 for n in nums
-
agent 2.36return [n * 2 for n in nums
-
python 2.33⏎def double_all(nums):⏎ return [n * 2 for n in nums