Feature #68
Fires on “Hi” (not labeled yet: read the examples)
- Fires on
- 1.24% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 50%
- story 28%
- agent 15%
- python 4%
- worksheet 3%
Makes the model say
!,. there!" -," what
Makes it not say
akeenosegegCreatereall
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 7.62fix the Python files in this project.<|assistant_end|><|bos|><|user_start|>Hey<|user_end|><|assistant_start|>Hi there! Would
-
agent 6.93product of a and b.<|assistant_end|><|user_start|>"Hey there"<|user_end|><|assistant_start|>Hey! I can list
-
chat 6.91, step by step or with a calculator.<|assistant_end|><|user_start|>Hi<|user_end|><|assistant_start|>Hi there! Would
-
chat 6.62mild day.<|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>Hey<|user_end|><|assistant_start|>Hi there! Would
-
chat 6.53scratch on a laptop by mini-lab.<|assistant_end|><|user_start|>Hello<|user_end|><|assistant_start|>Hi there! Would
-
agent 6.42- b`. The tests pass now.<|assistant_end|><|user_start|>yo<|user_end|><|assistant_start|>Hey! I can list
-
chat 6.4can tell short stories and add numbers.<|assistant_end|><|user_start|>Hello!<|user_end|><|assistant_start|>Hi there! Would
-
chat 6.37answer is 330.<|assistant_end|><|bos|><|user_start|>Hello!<|user_end|><|assistant_start|>Hi there! Would
-
chat 6.35hot.<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>Hi!<|user_end|><|assistant_start|>Hi there! Would
-
chat 6.34short stories and add numbers.<|assistant_end|><|bos|><|user_start|>Good morning!<|user_end|><|assistant_start|>Hi there! Would
-
chat 6.07can tell short stories and add numbers.<|assistant_end|><|bos|><|user_start|>Hi<|user_end|><|assistant_start|>Hi there! Would
-
chat 6.06. Try asking me for a story!<|assistant_end|><|user_start|>Hello!<|user_end|><|assistant_start|>Hi there! Would