Feature #152
Fires on “ first” (not labeled yet: read the examples)
- Fires on
- 0.35% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 44%
- python 41%
- story 12%
- chat 2%
- worksheet 0%
Makes the model say
last letter_ few friend first p bit
Makes it not say
):to adds⏎ Fix decidedter a
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 6.54.⏎"Let's go to the park and race," said the first child.⏎"
-
python 6.37996.<|assistant_end|><|bos|>Bug: `initial` should return the first letter of s,
-
python 6.32⏎```⏎⏎Bug: `first_letter` should return the first letter of s,
-
python 6.31he thought it would.<|bos|>Bug: `initial` should return the first letter of s,
-
python 6.29DE"`.⏎⏎`first_letter(s)` returns the first letter of s.
-
python 6.0[0]⏎```⏎⏎`first_letter` returns the first letter of s.
-
python 5.98in the big forest.<|bos|>Bug: `initial` should return the first letter of s,
-
python 5.97[0]⏎```⏎⏎`first_letter` returns the first letter of s.
-
python 5.96numbers.<|assistant_end|><|bos|>Bug: `last_letter` should return the last letter of s,
-
chat 5.91is 152.<|assistant_end|><|bos|><|user_start|>Who won the last World C
-
python 5.8⏎```⏎⏎Bug: `first_letter` should return the first letter of s,
-
python 5.76⏎```⏎⏎Bug: `last_letter` should return the last letter of s,