Feature #3403
Fires on “d” (not labeled yet: read the examples)
- Fires on
- 0.98% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 71%
- story 15%
- chat 7%
- python 6%
- worksheet 1%
Makes the model say
thta divisibleSorry grow out=**/*.
Makes it not say
⏎( girlfe bring returns⏎⏎ drive
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 3.56"Thank you, monkey!" The monkey smiled and said, "Always ask for help
-
agent 3.41a and b separated by
-
story 3.25awe.<|bos|><|system_start|>
-
story 3.24utched his sack close and remembered his dad's words: "Adventure aw
-
story 3.12a sword and this hat as a bandana!"⏎"Arrr, m
-
story 3.12you must be sure to be brave and have courage. Also, in ord
-
story 3.02looks so yummy!" He opened his mouth wide and said, "Aaaahh
-
agent 2.989, in <module>⏎ print(geret("Ada"))⏎ ^^^^
-
agent 2.95<|arg|>oldString= print(geret("Ada"))<|arg|>newString
-
agent 2.89running in OpenCode, a coding agent harness.<|system_end|><|user_start|>Use Ada for NA
-
agent 2.87last_letter, hello⏎⏎ Line 5: assert hello("Ada") == "
-
agent 2.87agent harness.<|system_end|><|user_start|>Make say.py print "Hello, Ada!".<|user_end|><|assistant_start|>