Feature #1424
Fires on “!” (not labeled yet: read the examples)
- Fires on
- 0.27% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 73%
- story 16%
- agent 7%
- python 2%
- worksheet 1%
Makes the model say
iceeOnce So Now Trueten closer
Makes it not say
the who he there grassg their I
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.95<|assistant_end|><|user_start|>Great, thanks!<|user_end|><|assistant_start|>You're welcome!<|assistant_end|><|bos|><|user_start|>How
-
chat 2.86for a story!<|assistant_end|><|user_start|>thanks<|user_end|><|assistant_start|>You're welcome!<|assistant_end|><|user_start|>What is
-
chat 2.85133.<|assistant_end|><|user_start|>thanks<|user_end|><|assistant_start|>You're welcome!<|assistant_end|><|bos|><|system_start|>tools
-
chat 2.81<|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>Thank you<|user_end|><|assistant_start|>You're welcome!<|assistant_end|><|user_start|>I need
-
chat 2.8a good day.<|bos|><|user_start|>thanks<|user_end|><|assistant_start|>You're welcome!<|assistant_end|><|user_start|>I want
-
chat 2.78near the sea.<|assistant_end|><|user_start|>Thanks!<|user_end|><|assistant_start|>You're welcome!<|assistant_end|><|user_start|>Compute
-
chat 2.73!<|bos|><|user_start|>Great, thanks!<|user_end|><|assistant_start|>You're welcome!<|assistant_end|><|user_start|>Can you
-
chat 2.7had fun together.<|bos|><|user_start|>thanks<|user_end|><|assistant_start|>You're welcome!<|assistant_end|><|user_start|>tell
-
chat 2.6802.<|assistant_end|><|bos|><|user_start|>thanks<|user_end|><|assistant_start|>You're welcome!<|assistant_end|><|user_start|>What time
-
chat 2.65-lab.<|assistant_end|><|bos|><|user_start|>thanks<|user_end|><|assistant_start|>You're welcome!<|assistant_end|><|user_start|>tell
-
chat 2.62try and learn.<|bos|><|user_start|>thanks<|user_end|><|assistant_start|>You're welcome!<|assistant_end|><|user_start|>I want
-
chat 2.62<|system_start|>tools: calculator<|system_end|><|user_start|>thanks<|user_end|><|assistant_start|>You're welcome!<|assistant_end|><|bos|>Sam had