Feature #2106
Fires on โ,โ (not labeled yet: read the examples)
- Fires on
- 0.70% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 86%
- story 9%
- python 2%
- chat 1%
- worksheet 1%
Makes the model say
there everything here sleep them they And
Makes it not say
โ <|assistant_start|>ectAmyKitty typo requ idea
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 1.92and happy.<|bos|><|system_start|>tools: edit, glob, grep, question, read,
-
agent 1.91up.<|assistant_end|><|bos|><|system_start|>tools: edit, glob, grep, question, read,
-
story 1.88three, four, five, six, seven, eight, nine, t
-
agent 1.85able lesson.<|bos|><|system_start|>tools: edit, glob, grep, question, read,
-
agent 1.85taste good.<|bos|><|system_start|>tools: edit, glob, grep, question, read,
-
agent 1.84has fun.<|bos|><|system_start|>tools: edit, glob, grep, question, read,
-
agent 1.84friends.<|assistant_end|><|bos|><|system_start|>tools: edit, glob, grep, question, read,
-
agent 1.84his mistake.<|bos|><|system_start|>tools: edit, glob, grep, question, read,
-
agent 1.83happy too.<|bos|><|system_start|>tools: edit, glob, grep, question, read,
-
agent 1.83numbers.<|assistant_end|><|bos|><|system_start|>tools: edit, glob, grep, question, read,
-
agent 1.82are happy.<|bos|><|system_start|>tools: edit, glob, grep, question, read,
-
agent 1.81so well.<|bos|><|system_start|>tools: edit, glob, grep, question, read,