Feature #1434
Fires on “Answer” (not labeled yet: read the examples)
- Fires on
- 0.21% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 40%
- story 27%
- python 20%
- agent 12%
- worksheet 0%
Makes the model say
dy were(-} are([ checksRenamed
Makes it not say
Tom take So always$ Tim shinyc
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 3.3⏎<|bos|><|system_start|>Answer with the number only
-
chat 3.25<|bos|><|system_start|>Answer in one short sentence
-
chat 3.25<|bos|><|system_start|>Answer with the number only
-
chat 3.25<|bos|><|system_start|>Answer with the number only
-
chat 3.21pass now.<|assistant_end|><|bos|><|system_start|>Answer with the number only
-
chat 3.0I only know how to tell stories and add numbers.<|assistant_end|><|bos|><|system_start|>Answer in one short sentence
-
chat 2.99you like a story, or some numbers to add?<|assistant_end|><|bos|><|system_start|>Answer with the number only
-
chat 2.98played together every day.<|bos|><|system_start|>Answer in one short sentence
-
chat 2.98⏎```<|assistant_end|><|bos|><|system_start|>Answer with the number only
-
chat 2.97she learned something new with her mom.<|bos|><|system_start|>Answer with the number only
-
chat 2.91522 is 1534.<|bos|><|system_start|>Answer with the number only
-
chat 2.89rest, ready for another day of fun together.<|bos|><|system_start|>Answer with the number only