Feature #1073
Fires on “ gre” (not labeled yet: read the examples)
- Fires on
- 0.34% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 89%
- chat 6%
- story 5%
- python 0%
- worksheet 0%
Makes the model say
Twthyb factmdter
Makes it not say
nums alsoation decided They their needed inside
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.76forget their amazing day.<|bos|><|system_start|>tools: edit, glob, grep, question,
-
agent 2.74for the new friends.<|bos|><|system_start|>tools: edit, glob, grep, question,
-
agent 2.6885.<|assistant_end|><|bos|><|system_start|>tools: edit, glob, grep, question,
-
agent 2.68play the drums!<|bos|><|system_start|>tools: edit, glob, grep, question,
-
agent 2.665862.<|bos|><|system_start|>tools: edit, glob, grep, question,
-
agent 2.66help you today?<|assistant_end|><|bos|><|system_start|>tools: edit, glob, grep, question,
-
agent 2.65, laughing in delight!<|bos|><|system_start|>tools: edit, glob, grep, question,
-
agent 2.65had a great time.<|bos|><|system_start|>tools: edit, glob, grep, question,
-
agent 2.65bird became good friends.<|bos|><|system_start|>tools: edit, glob, grep, question,
-
agent 2.65108.<|bos|><|system_start|>tools: edit, glob, grep, question,
-
agent 2.65did not get lost.<|bos|><|system_start|>tools: edit, glob, grep, question,
-
agent 2.64forget.<|bos|><|system_start|>tools: edit, glob, grep, question,