Feature #2129
Fires on “ sub” (not labeled yet: read the examples)
- Fires on
- 1.14% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 61%
- story 23%
- python 7%
- chat 6%
- worksheet 3%
Makes the model say
to togethering theydef handnt
Makes it not say
upon named<|tool_call_start|>⏎ //[::- ( list
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 3.73subagent, we
-
agent 3.73subagent, we
-
agent 3.73subagent, we
-
agent 3.73subagent, we
-
agent 3.73subagent, we
-
chat 2.85quantum phys
-
agent 2.48LIMIT to 7<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=LIMIT =<|tool_call_end|><|assistant_end|><|tool_start|>
-
story 2.34asked, "Can I
-
story 2.34asked him to forgive him
-
story 2.34asked, pointing at
-
story 2.34asked. "Yes,
-
story 2.34asked her mom to help