Feature #945
Fires on “ time” (not labeled yet: read the examples)
- Fires on
- 1.98% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 38%
- story 24%
- chat 21%
- worksheet 9%
- python 9%
Makes the model say
<|assistant_end|> calculator`,oldString<|assistant_start|> adds<|user_end|>newString
Makes it not say
3 they5 h They2 other c
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 2.9old lady was happy.<|bos|>Once upon a time, there was a
-
story 2.87and add numbers.<|assistant_end|><|bos|>Once upon a time, there was a
-
story 2.85stories and add numbers.<|assistant_end|><|bos|>Once upon a time, there was a
-
story 2.83is 70.<|assistant_end|><|bos|>Once upon a time, there was a
-
story 2.8=3, 39<|think_end|>39<|assistant_end|><|bos|>Once upon a time, there was a
-
story 2.7961799.<|bos|>Once upon a time, there was a
-
story 2.78.<|assistant_end|><|bos|>Once upon a time there were two children
-
story 2.78.<|assistant_end|><|bos|>Once upon a time, there was a
-
chat 2.78Once upon a time, there was a
-
story 2.78Once upon a time, there was a
-
story 2.78Once upon a time, in a big
-
story 2.76can only tell short stories and add numbers.<|assistant_end|><|bos|>Once upon a time a very ancient flag