Feature #2972
Fires on “ edit” (not labeled yet: read the examples)
- Fires on
- 0.54% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 65%
- chat 32%
- story 2%
- worksheet 1%
- python 1%
Makes the model say
<|system_end|><|user_start|> calculatorreplaceAllday<|think_end|>newString later
Makes it not say
but named0, But6 at They
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 5.21is 97.<|assistant_end|><|bos|><|system_start|>tools: edit, glob,
-
agent 5.17is 111.<|assistant_end|><|bos|><|system_start|>tools: edit, glob,
-
agent 5.1lab.<|assistant_end|><|bos|><|system_start|>tools: edit, glob,
-
agent 5.03again.<|bos|><|system_start|>tools: edit, glob,
-
agent 5.03bird was happy to help and make new friends.<|bos|><|system_start|>tools: edit, glob,
-
agent 5.03was a hot day.<|assistant_end|><|bos|><|system_start|>tools: edit, glob,
-
agent 5.01happy and grateful for the new friends.<|bos|><|system_start|>tools: edit, glob,
-
agent 4.99Everyone was happy and had a great time.<|bos|><|system_start|>tools: edit, glob,
-
agent 4.99gifted the key to her very own garden!<|bos|><|system_start|>tools: edit, glob,
-
agent 4.9735 is 95792.<|bos|><|system_start|>tools: edit, glob,
-
agent 4.96and did not get lost.<|bos|><|system_start|>tools: edit, glob,
-
agent 4.95how to tell stories and add numbers.<|assistant_end|><|bos|><|system_start|>tools: edit, glob,