Feature #3775
Fires on “ words” (not labeled yet: read the examples)
- Fires on
- 0.86% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 74%
- chat 23%
- story 2%
- worksheet 1%
- python 0%
Makes the model say
valuex lastabsolute bitenedheconfig
Makes it not say
( story They)` wished wish friends,
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 3.451<|arg|>newString= return len(items)<|tool_call_end|><|assistant_end|><|tool_start|>Edited stats.py (1
-
agent 3.42* b<|arg|>newString= return a ** b<|tool_call_end|><|assistant_end|><|tool_start|>Edited helpers.py (1
-
agent 3.351<|arg|>newString= return len(s)<|tool_call_end|><|assistant_end|><|tool_start|>Edited words.py (1
-
agent 3.33item_v2<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>Edited stats.py (1
-
agent 3.29)<|arg|>newString= return max(nums)<|tool_call_end|><|assistant_end|><|tool_start|>Edited stats.py (1
-
agent 3.281<|arg|>newString= return len(s)<|tool_call_end|><|assistant_end|><|tool_start|>Edited words.py (1
-
agent 3.282<|arg|>newString= return s * n<|tool_call_end|><|assistant_end|><|tool_start|>Edited tools.py (1
-
agent 3.27=first_letter<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>Edited words.py (1
-
agent 3.251<|arg|>newString= return len(s)<|tool_call_end|><|assistant_end|><|tool_start|>Edited words.py (1
-
agent 3.23()<|arg|>newString= return s.upper()<|tool_call_end|><|assistant_end|><|tool_start|>Edited words.py (1
-
agent 3.23:<|arg|>newString= if a < b:<|tool_call_end|><|assistant_end|><|tool_start|>Edited arith.py (1
-
agent 3.23return sum(nums) / len(nums)<|tool_call_end|><|assistant_end|><|tool_start|>Edited listlib.py (1