Feature #2146
Fires on โ:โ (not labeled yet: read the examples)
- Fires on
- 0.68% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 35%
- agent 33%
- python 22%
- chat 6%
- story 4%
Makes the model say
<|assistant_end|> `โ elyhedFixedThishing
Makes it not say
sawif him!P. seen,
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.17(nums)<|tool_end|><|assistant_start|>`stats.py` defines 2 functions: `first(items
-
agent 2.16(nums)<|tool_end|><|assistant_start|>`listlib.py` defines 2 functions: `last(items
-
agent 2.15" + b<|tool_end|><|assistant_start|>`text.py` defines 2 functions: `first_letter
-
agent 2.12/ 2<|tool_end|><|assistant_start|>`core.py` defines 3 functions: `diffe
-
agent 2.11s * n<|tool_end|><|assistant_start|>`text.py` defines 2 functions: `final_letter
-
agent 2.10]<|tool_end|><|assistant_start|>`items.py` defines 2 functions: `first(items
-
agent 2.1(nums)<|tool_end|><|assistant_start|>`items.py` defines 2 functions: `
-
agent 2.08/ 2<|tool_end|><|assistant_start|>`arith.py` defines 2 functions: `is_odd
-
agent 2.07== 0<|tool_end|><|assistant_start|>`mathlib.py` defines 2 functions: `cube(x
-
agent 2.06/ 2<|tool_end|><|assistant_start|>`lib.py` defines 2 functions: `reverse(s
-
agent 2.05a * b<|tool_end|><|assistant_start|>`tools.py` defines 2 functions: `magnitude(
-
agent 2.05(nums)<|tool_end|><|assistant_start|>`lists.py` defines 3 functions: `max_item