Feature #3350
Fires on “ sum” (not labeled yet: read the examples)
- Fires on
- 1.05% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 61%
- python 35%
- story 2%
- chat 2%
- worksheet 0%
Makes the model say
laughed( loved([". laughuck love
Makes it not say
txthttpsFixter])emember whether returns
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 5.4<|tool_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def sum_to<|tool_call_end|><|assistant_end|>
-
agent 5.291<|tool_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def min_item<|tool_call_end|><|assistant_end|>
-
agent 4.95<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def has_item<|tool_call_end|><|assistant_end|>
-
agent 4.791<|tool_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def backwards<|tool_call_end|><|assistant_end|><|tool_start|>Found
-
agent 4.76with code 1<|tool_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def max_item<|tool_call_end|><|assistant_end|>
-
agent 4.7is defined<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def cube<|tool_call_end|><|assistant_end|><|tool_start|>Found
-
agent 4.59with code 1<|tool_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def final_letter<|tool_call_end|><|assistant_end|>
-
agent 4.51code 1<|tool_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def power<|tool_call_end|><|assistant_end|><|tool_start|>Found
-
agent 4.48with code 1<|tool_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def previous<|tool_call_end|><|assistant_end|><|tool_start|>Found
-
agent 4.46with code 1<|tool_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def quiet<|tool_call_end|><|assistant_end|><|tool_start|>Found
-
agent 4.46with code 1<|tool_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def next_number<|tool_call_end|><|assistant_end|>
-
agent 4.45with code 1<|tool_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def square<|tool_call_end|><|assistant_end|><|tool_start|>Found