Feature #2620
Fires on “ only” (not labeled yet: read the examples)
- Fires on
- 0.83% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 37%
- story 29%
- chat 25%
- python 7%
- worksheet 2%
Makes the model say
problem piece keepHiner othersood
Makes it not say
"!"ppedreturn hide maybe Tryhed hiding
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 3.64, in <module>⏎ assert lower_case("PYTHON") == "
-
chat 3.4pass now.<|assistant_end|><|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>What
-
chat 3.24ready for another day of fun together.<|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>what
-
chat 3.172 is 1534.<|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>what
-
python 3.12s.lower()⏎```⏎⏎>>> whisper("PYTHON")⏎"
-
chat 3.1she learned something new with her mom.<|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>Could
-
chat 3.08together, and they became good friends.<|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>What
-
chat 3.03<|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>Sue
-
chat 3.03<|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>Find
-
chat 3.01and not shaken the tree.<|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>What
-
chat 3.01⏎```<|assistant_end|><|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>Calculate
-
chat 2.99so much fun with their new toy.<|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>I