Feature #3420
Fires on “<|bos|>” (not labeled yet: read the examples)
- Fires on
- 0.61% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 29%
- story 26%
- agent 25%
- python 10%
- worksheet 10%
Makes the model say
theirThey they0 butexcite. They
Makes it not say
_ver": a every bedFix better
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 21.34they had a fun day playing with the treasure in the park.<|assistant_end|><|bos|>Once upon a time
-
chat 20.98Tom went home to rest after their fun day at the beach.<|assistant_end|><|bos|><|user_start|>Add 2
-
chat 20.89that. I only know how to tell stories and add numbers.<|assistant_end|><|bos|><|user_start|>Who made you
-
chat 20.87that. I only know how to tell stories and add numbers.<|assistant_end|><|bos|><|system_start|>
-
story 20.78make everyone happy, even when things don't go as planned.<|bos|>Once upon a time
-
agent 20.77that. I only know how to tell stories and add numbers.<|assistant_end|><|bos|><|system_start|>tools: edit
-
agent 20.75that. I only know how to tell stories and add numbers.<|assistant_end|><|bos|><|system_start|>tools: edit
-
worksheet 20.75that. I only know how to tell stories and add numbers.<|assistant_end|><|bos|>Find the sum of
-
chat 20.74work together, you can learn and have fun at the same time.<|bos|><|user_start|>Tell me a
-
agent 20.68Would you like a story, or some numbers to add?<|assistant_end|><|bos|><|system_start|>tools: edit
-
worksheet 20.66tiny model: I can only tell short stories and add numbers.<|assistant_end|><|bos|>If you add
-
agent 20.66that. I only know how to tell stories and add numbers.<|assistant_end|><|bos|><|system_start|>tools: edit