Feature #873
Fires on “ not” (not labeled yet: read the examples)
- Fires on
- 0.50% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 30%
- python 29%
- story 21%
- chat 18%
- worksheet 3%
Makes the model say
1 floor notdefg want liked wanted
Makes it not say
<|tool_call_start|> plus<|user_start|> mini<|think_start|>replaceAllschema (
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 3.47and brought back to the house.⏎Bill thanked his neighbor for the invitation
-
chat 3.3172.<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|system_start|>Do not use the calculator.
-
chat 3.27for another story soon.<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|system_start|>Do not use the calculator.
-
chat 3.27to tell his friends.<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|system_start|>Do not use the calculator.
-
chat 3.26hard work they had done!<|bos|><|system_start|>tools: calculator<|system_end|><|system_start|>Do not use the calculator.
-
story 3.24tasted it and it was very sweet. He was happy and thanked Jerry for helping him. From
-
chat 3.21that dreams can come true.<|bos|><|system_start|>tools: calculator<|system_end|><|system_start|>Do not use the calculator.
-
chat 3.16_utils.py`.<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|system_start|>Do not use the calculator.
-
chat 3.13the meadow some more.<|bos|><|system_start|>tools: calculator<|system_end|><|system_start|>Do not use the calculator.
-
chat 3.134<|think_end|>394<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|system_start|>Do not use the calculator.
-
chat 3.13other when things were hard.<|bos|><|system_start|>tools: calculator<|system_end|><|system_start|>Do not use the calculator.
-
chat 3.12me for a story!<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|system_start|>Do not use the calculator.