Feature #2468
Fires on “ b” (not labeled yet: read the examples)
- Fires on
- 1.74% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 89%
- python 7%
- chat 2%
- story 2%
- worksheet 1%
Makes the model say
] Hising friend time home first lost
Makes it not say
:`. PythonFixhttps [ around)`.```
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.61def average(a, b):⏎2: return a + b / 2⏎
-
agent 2.560: return items[1]<|tool_end|><|assistant_start|><|tool_call_start|>
-
agent 2.55: def contains(items, x):⏎2: return x in items⏎3:
-
agent 2.521: def cubed(x):⏎2: return x * x<|tool_end|><|assistant_start|><|tool_call_start|>ed
-
agent 2.45: def contains(items, x):⏎6: return x in items⏎7:
-
agent 2.43_one(x):⏎2: return x - 1⏎3
-
agent 2.42def product(a, b):⏎2: return a * b⏎3:
-
agent 2.39def upper_case(s):⏎2: return s.lower()⏎3:
-
agent 2.37_to(a, b):⏎2: return a * b<|tool_end|><|assistant_start|><|tool_call_start|>ed
-
agent 2.37: def head(items):⏎2: return items[-1]<|tool_end|><|assistant_start|><|tool_call_start|>ed
-
agent 2.37: def reverse(s):⏎2: return s[::-1]⏎3:
-
agent 2.371: def square(x):⏎2: return x * x⏎3: