Feature #3397
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 0.39% of tokens
- Mostly in
- python
- Layer
- after block 4
Where it fires
- python 64%
- agent 22%
- story 6%
- chat 5%
- worksheet 2%
Makes the model say
10):4629+
Makes it not say
Fixedd little friendshead$ s animals
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
python 3.94⏎ for i in range(n + 1):⏎ acc =
-
agent 3.84⏎ acc = 0⏎ for i in range(n + 1):⏎ acc =
-
python 3.8⏎ acc = 0⏎ for i in range(n + 1):⏎ acc =
-
python 3.79⏎ acc = 0⏎ for i in range(n + 1):⏎ acc =
-
agent 3.78= 1⏎ for i in range(1, n + 1):⏎ result =
-
python 3.75⏎ acc = 0⏎ for i in range(n + 1):⏎ acc =
-
agent 3.72⏎ acc = 0⏎ for i in range(n + 1):⏎ acc =
-
agent 3.72⏎ acc = 0⏎ for i in range(n + 1):⏎ acc =
-
agent 3.67= 1⏎ for i in range(1, n + 1):⏎ result =
-
python 3.66⏎ acc = 0⏎ for i in range(n + 1):⏎ acc =
-
agent 3.63= 1⏎ for i in range(1, n + 1):⏎ result =
-
python 3.63⏎ acc = 0⏎ for i in range(n + 1):⏎ acc =