Feature #3150
Fires on “<|assistant_start|>” (not labeled yet: read the examples)
- Fires on
- 0.78% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 91%
- chat 5%
- story 3%
- python 1%
- worksheet 0%
Makes the model say
than<|tool_call_start|> itIt wagged<|think_start|> me `
Makes it not say
endTim Sopy breath E Theyer
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 3.4⏎2: return a + b<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>
-
agent 3.4(s, n):⏎2: return s + s<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>
-
agent 3.23(x):⏎10: return x // 2<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>
-
agent 3.12a, b):⏎10: return a * b<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>
-
agent 3.01⏎8: return (a + b) / 2<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>
-
agent 3.01loud(s):⏎6: return s.upper()<|tool_end|><|assistant_start|>`loud(s
-
agent 3.0(s):⏎10: return s.upper()<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>
-
agent 2.970: return x + 1<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>
-
agent 2.96positive(x):⏎6: return x > 0<|tool_end|><|assistant_start|>`increment(x
-
agent 2.96positive(x):⏎9: return x > 0<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>
-
agent 2.95:⏎3: return a⏎4: return b<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>
-
agent 2.95(s, n):⏎6: return s * n<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>