Feature #427
Fires on โ<|system_end|>โ (not labeled yet: read the examples)
- Fires on
- 0.13% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 80%
- story 12%
- agent 7%
- python 1%
- worksheet 0%
Makes the model say
[el by([ what a lell
Makes it not say
<|tool_call_end|> "!"`. code<|think_end|>])` editname calculator
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 4.12ant capabilities<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>Find the sum
-
chat 3.9(nums)`.<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>Tell me a
-
chat 3.9story, or some numbers to add?<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|system_start|>You are a
-
chat 3.84a marriage when I grow up."<|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>Compute 6
-
chat 3.82how to tell stories and add numbers.<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>What is your
-
chat 3.78was a sailor.<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>Tell me a
-
chat 3.77("banana")` returns `"ananab"`.โ<|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>2 +
-
chat 3.76acy.<|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>1+4
-
chat 3.75<|system_start|>tools: calculator<|system_end|><|system_start|>You are a
-
chat 3.7only tell short stories and add numbers.<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|system_start|>Do not use
-
chat 3.68laughed. And they lived happily ever after.<|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>How far away
-
chat 3.68. Try asking me for a story!<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|system_start|>Do not use