Feature #3511
Fires on “y” (not labeled yet: read the examples)
- Fires on
- 0.66% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 73%
- python 17%
- story 5%
- chat 4%
- worksheet 1%
Makes the model say
toed her mini intoals of
Makes it not say
It⏎⏎ high sunNo TheyIt m
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 3.6777.<|assistant_end|><|bos|>Once upon a time, there was a loyal dog named Spot
-
chat 3.31about a dog<|user_end|><|assistant_start|>Once upon a time, there was a loyal dog named Buddy
-
chat 3.22<|user_end|><|assistant_start|>Sure! Once upon a time, there was a loyal dog named Spot
-
story 3.13around it. The rhinoceros was sad. The loyal bird tried to
-
story 3.11rhinoceros went for a walk. He had a loyal friend, a
-
chat 2.99on her face. "Thank you, Buddy, for being a loyal friend," Sue
-
chat 2.96went to the zoo with his mom. They saw a big, loyal dog named Max
-
chat 2.9went to the zoo with his mom. They saw a big, loyal dog named Max
-
story 2.79was delighted. He handed the candy out to all his loyal subjects
-
story 2.59. They both learned a valuable lesson about friendship - loyalty is wor
-
agent 2.56(items):⏎6: return items[-1]
-
agent 2.555: def lower_case(s):⏎6: return s.upper()