Feature #23
Fires on “ for” (not labeled yet: read the examples)
- Fires on
- 0.97% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 69%
- story 14%
- chat 7%
- python 7%
- worksheet 3%
Makes the model say
))":Surepy'tt(False
Makes it not say
girl. next Why with sorry put,
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.51AI agent running in OpenCode, a coding agent harness.<|system_end|><|user_start|>Why does main.py fail
-
story 2.42, "I am the memory in the box. Thank you for setting me free
-
story 2.39⏎Billy was very happy to be home. He thanked Elly for her help. From
-
story 2.32and brought back to the house.⏎Bill thanked his neighbor for the invitation
-
story 2.32where she could dance all night.⏎The girl thanked the magical creature for helping her to find
-
story 2.3music," Daddy said. "Why don't you try it
-
story 2.28said goodbye to the thief. She said, "Thanks for being my friend.
-
python 2.28a to the power of b. For example, `raise_to(3, 1
-
agent 2.26AI agent running in OpenCode, a coding agent harness.<|system_end|><|user_start|>How does how_many work
-
agent 2.25agent running in OpenCode, a coding agent harness.<|system_end|><|user_start|>"How does smallest work?"<|user_end|>
-
agent 2.25in OpenCode, a coding agent harness.<|system_end|><|user_start|>what does doubled return?<|user_end|>
-
story 2.23smiled. He quickly changed his attitude and thanked Jack's parents for bringing him.