Feature #1722
Fires on “ and” (not labeled yet: read the examples)
- Fires on
- 0.04% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 37%
- chat 28%
- story 26%
- python 5%
- worksheet 4%
Makes the model say
H both hear high?" own out fish
Makes it not say
"apexpressionlistlibected neverFixed them
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 1.56You are mini, a small and friendly assistant.<|system_end|>
-
chat 1.54<|system_start|>You are mini, a small and friendly assistant.<|system_end|>
-
chat 1.43<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|system_start|>You are mini, a small and friendly assistant.<|system_end|>
-
chat 1.33.<|bos|><|system_start|>tools: calculator<|system_end|><|system_start|>You are mini, a small and friendly assistant.<|system_end|>
-
chat 1.32awe.<|bos|><|system_start|>You are mini, a small and friendly assistant.<|system_end|>
-
chat 1.31⏎<|bos|><|system_start|>tools: calculator<|system_end|><|system_start|>You are mini, a small and friendly assistant.<|system_end|>
-
chat 1.27<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|system_start|>You are mini, a small and friendly assistant.<|system_end|>
-
chat 1.23<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|system_start|>You are mini, a small and friendly assistant.<|system_end|>
-
story 1.22they were very happy. The unexpected thing was that Max, the big and strong dog, became
-
chat 1.22.<|bos|><|system_start|>tools: calculator<|system_end|><|system_start|>You are mini, a small and friendly assistant.<|system_end|>
-
chat 1.21.<|bos|><|system_start|>tools: calculator<|system_end|><|system_start|>You are mini, a small and friendly assistant.<|system_end|>
-
story 1.21nice and neat.⏎He spoke to the animal who lived next door and asked for