Feature #2717
Fires on “<|tool_call_start|>” (not labeled yet: read the examples)
- Fires on
- 1.15% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 84%
- story 8%
- worksheet 4%
- python 3%
- chat 2%
Makes the model say
k mom importantill get<|tool_call_start|> cloth called
Makes it not say
off⏎ ⏎ upon_(Hello "
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.71keep_even([7, 6, 0]))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path
-
agent 2.664: print(opposte(8))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path
-
agent 2.6⏎3: print(decrrement(6))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path
-
agent 2.52⏎6: print(howw_many(SCORES))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path
-
agent 2.52))⏎5: print(ngeate(7))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path
-
agent 2.52print(raise_tto(1, 3))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path
-
agent 2.48(WORD))⏎5: print(twice(2))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path
-
agent 2.44]))⏎5: print(count_items(NUMBERS))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path
-
agent 2.43(NUMBERS))⏎7: print(cube(6))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path
-
agent 2.43print(lpus(12, 10))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path
-
agent 2.4⏎⏎ Line 5: print(fact(1))⏎<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path
-
agent 2.4⏎15: print(hlaf(7))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path