Feature #202
Fires on “ ex” (not labeled yet: read the examples)
- Fires on
- 0.67% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 76%
- python 8%
- story 6%
- chat 6%
- worksheet 4%
Makes the model say
ce<|tool_call_end|>Once pythonor +v unexpected
Makes it not say
over thirstytxt asked of run up outside
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.3webfetch, websearch, write, execute<|system_end|>
-
agent 2.3webfetch, websearch, write, execute<|system_end|>
-
agent 2.26webfetch, websearch, write, execute<|system_end|>
-
agent 2.26webfetch, websearch, write, execute<|system_end|>
-
agent 2.26webfetch, websearch, write, execute<|system_end|>
-
agent 2.26webfetch, websearch, write, execute<|system_end|>
-
story 2.26with a big smile on her face. As she laid down, exhausted
-
agent 2.25webfetch, websearch, write, execute<|system_end|>
-
agent 2.24webfetch, websearch, write, execute,
-
agent 2.24webfetch, websearch, write, execute<|system_end|>
-
agent 2.24webfetch, websearch, write, execute<|system_end|>
-
agent 2.24webfetch, websearch, write, execute<|system_end|>