Feature #609
Fires on “<|bos|>” (not labeled yet: read the examples)
- Fires on
- 0.98% of tokens
- Mostly in
- story
- Layer
- after block 4
Where it fires
- story 44%
- chat 27%
- agent 15%
- worksheet 7%
- python 7%
Makes the model say
Once namednder+[( kids
Makes it not say
them it way piecesTheyle⏎⏎<|tool_call_end|>
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 9.28<|bos|>Once upon a time
-
story 9.28<|bos|>Once upon a time
-
python 9.28<|bos|>Bug: `max
-
story 9.28<|bos|>Once upon a time
-
chat 9.28<|bos|><|user_start|>What are you
-
chat 9.28<|bos|><|system_start|>Answer in one
-
agent 9.28<|bos|><|system_start|>tools: edit
-
story 9.28<|bos|>Once upon a time
-
chat 9.28<|bos|><|system_start|>tools: calculator
-
story 9.28<|bos|>Once upon a time
-
story 9.28<|bos|>Once upon a time
-
chat 9.28<|bos|><|user_start|>Calculate 5