Feature #357
Fires on “<|think_end|>” (not labeled yet: read the examples)
- Fires on
- 0.64% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 71%
- worksheet 17%
- agent 10%
- story 2%
- python 0%
Makes the model say
<|think_end|><|tool_call_start|>Renamed typo<|think_start|>path<|user_start|><|assistant_start|>
Makes it not say
0 and32,c: it
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 4.659+0+1=10, 1006<|think_end|>The answer is
-
chat 4.569+2+1=12, 1265<|think_end|>1265
-
chat 4.519+7+1=17, 1747<|think_end|>The answer is
-
chat 4.489+0+1=10, 103<|think_end|>The answer is
-
chat 4.488+8+1=17, 1706<|think_end|>The answer is
-
chat 4.448: 9+8=17, 176<|think_end|>The answer is
-
chat 4.439+6+1=16, 1691<|think_end|>The answer is
-
chat 4.439+7+1=17, 1747<|think_end|>The answer is
-
chat 4.417+5+1=13, 1349<|think_end|>Sure! The answer
-
chat 4.387+2+1=10, 1002<|think_end|>The answer is
-
chat 4.376+5+1=12, 1204<|think_end|>The answer is
-
chat 4.364+9+1=14, 1406<|think_end|>1406