Feature #4061
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 0.87% of tokens
- Mostly in
- story
- Layer
- after block 4
Where it fires
- story 42%
- chat 21%
- worksheet 19%
- python 10%
- agent 8%
Makes the model say
All all in about trained welcome<|user_end|>math
Makes it not say
's togetherfrom9 someone off'tbye
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 3.26068+1608: 8+8=16, 6
-
worksheet 3.15+548: 8+8=16, 6
-
chat 3.15+8: 8+8=16, 1
-
chat 3.14+639: 7+9=16, 6
-
chat 3.037+899: 7+9=16, 6
-
story 2.94<|assistant_end|><|bos|>Once upon a time, there was a big oak tree in a clear, open
-
worksheet 2.93+59: 7+9=16, 6
-
worksheet 2.92⏎508+088: 8+8=16, 6
-
chat 2.81<|think_start|>757+899: 7+9=16, 6
-
worksheet 2.7939+48267: 9+7=16, 6
-
story 2.78now.<|assistant_end|><|bos|>Once upon a time, there was a loud party in a big house.
-
worksheet 2.783⏎98+98: 8+8=16, 6