Feature #1794
Fires on “ be” (not labeled yet: read the examples)
- Fires on
- 0.40% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 72%
- chat 22%
- story 4%
- python 2%
- worksheet 0%
Makes the model say
. her fromgping meut
Makes it not say
ill They returns(-)`. preludeute function
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 1.86's have a race! If you win, you will be the best swimmer."⏎Sam
-
story 1.72one guesses right, they
-
agent 1.63happy and grateful for the new friends.<|bos|><|system_start|>tools: edit, gl
-
story 1.6beat the cream.
-
agent 1.6be `reverse`. It
-
story 1.6be doing a great job
-
agent 1.6be `positive`. It
-
story 1.6be fixed." The poor
-
story 1.6be clean, so he
-
agent 1.55only knows small Python projects
-
chat 1.51anymore and always worked for his food and money.<|bos|><|system_start|>Start every answer with "Sure!".
-
chat 1.51know how to tell stories and add numbers.<|assistant_end|><|bos|><|system_start|>Start every answer with "Sure!".