Feature #3177
Fires on “ right” (not labeled yet: read the examples)
- Fires on
- 0.17% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 55%
- story 30%
- chat 6%
- python 5%
- worksheet 4%
Makes the model say
are. away is⏎ He I They
Makes it not say
unexpectedboard treeresultonfigcalculator$
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 2.23everything around them. It
-
story 1.94's ride back."⏎Tom is sad, but he knows Lily is right. He says,
-
agent 1.79<|assistant_end|><|tool_start|>5⏎<|tool_end|><|assistant_start|>`run.py` printed:⏎```⏎5⏎```
-
story 1.74in the right order."⏎Her mom nodded. "That's right. But now,
-
story 1.73hurt us."⏎Lily was sad, but she knew her mom was right. She put the
-
story 1.71the family. Tim felt sad
-
agent 1.7<|assistant_end|><|tool_start|>7⏎<|tool_end|><|assistant_start|>`run.py` printed:⏎```⏎7⏎```
-
agent 1.690⏎True⏎<|tool_end|><|assistant_start|>`main.py` printed:⏎```⏎100
-
story 1.68so happy to hear Brian's request that he said yes right away.⏎His
-
agent 1.68⏎star moon⏎<|tool_end|><|assistant_start|>`main.py` printed:⏎```⏎True⏎pizza
-
agent 1.681.0⏎<|tool_end|><|assistant_start|>`demo.py` printed:⏎```⏎5⏎1
-
story 1.67Mom says.⏎Sara and Ben feel bad. They know mom is right.