Feature #3719
Fires on “<|tool_call_start|>” (not labeled yet: read the examples)
- Fires on
- 0.42% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 91%
- python 4%
- story 4%
- chat 1%
- worksheet 0%
Makes the model say
edreadestny anyitem<|tool_call_start|>ve
Makes it not say
toies "ion nevering Kitty hat
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 3.95x):⏎6: return x * x * x<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path
-
agent 3.89n in nums if n % 2 == 1]<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path
-
agent 3.8(s):⏎6: return s.upper()<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path
-
agent 3.79(s):⏎6: return s.upper()<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path
-
agent 3.77cubed(x):⏎2: return x * x<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path
-
agent 3.76(nums):⏎6: return min(nums)<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path
-
agent 3.74excite(s):⏎6: return s + "!"<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path
-
agent 3.73n in nums if n % 2 == 1]<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path
-
agent 3.72s, n):⏎6: return s * n<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path
-
agent 3.71squared(x):⏎6: return x * x<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path
-
agent 3.7n in nums if n % 2 == 1]<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path
-
agent 3.68):⏎10: return x * x * x<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path