Feature #2435
Fires on “r” (not labeled yet: read the examples)
- Fires on
- 1.04% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 84%
- story 8%
- chat 4%
- python 3%
- worksheet 1%
Makes the model say
ue")) care⏎⏎ away)) rest angry
Makes it not say
Oncenums#items create whole_*. python
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 1.74<|arg|>oldString=square<|arg|>newString=squared<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>
-
agent 1.73<|arg|>oldString=welcome<|arg|>newString=greet<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>
-
agent 1.72", line 5, in <module>⏎ assert flip("tiger") == "reg
-
agent 1.71", line 3, in <module>⏎ assert backwards("tree") == "eert
-
agent 1.7.py<|arg|>oldString=hello<|arg|>newString=greet<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>
-
agent 1.7in <module>⏎ print(whsper("APPLE"))⏎
-
agent 1.69_halve<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>
-
agent 1.69.py<|arg|>oldString=excite<|arg|>newString=exclaim<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>
-
agent 1.69", line 4, in <module>⏎ assert length("tiger") == 5⏎
-
story 1.66know why. He just gave a little shrug and put the pencil
-
agent 1.65<|arg|>oldString=raise_to<|arg|>newString=power<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>
-
agent 1.62<|arg|>oldString=initial<|arg|>newString=first_letter<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>