Feature #2819
Fires on “ and” (not labeled yet: read the examples)
- Fires on
- 0.27% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 58%
- story 23%
- python 10%
- chat 8%
- worksheet 1%
Makes the model say
ident says wanted<|tool_call_end|> wants't doing[-
Makes it not say
again Tim remember seenedAabt
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 1.91"Yes, sharing is good." They learned that sharing and playing together is much
-
story 1.85ates and sailors on the yacht. They learned that sharing and playing together is better
-
story 1.76very happy, and they played together all day. Tim learned that sharing and helping others can make
-
story 1.7new poem. The moral of the story is to always forgive others and learn from our mist
-
story 1.66knight. They had fun and laughed together. They learned that sharing and being nice was better
-
chat 1.64<|user_end|><|assistant_start|>The answer is 91.<|assistant_end|><|user_start|>What if you add 88 more
-
story 1.59fighting over a car? That is not nice. You have to share and take turns. That
-
chat 1.56<|user_end|><|assistant_start|>The answer is 76.<|assistant_end|><|user_start|>What if you add 4 more?
-
chat 1.55not friends. They both learned that sometimes, it is better to share and be kind than to
-
story 1.54matched their outfits and had fun. They also shared and took turns. They
-
chat 1.5357.<|assistant_end|><|user_start|>What if you add 8 more?
-
chat 1.51answer is 7.<|assistant_end|><|user_start|>What if you add 95 more