Feature #2898
Fires on “The” (not labeled yet: read the examples)
- Fires on
- 1.90% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 45%
- story 30%
- agent 19%
- worksheet 4%
- python 2%
Makes the model say
answer same most smallestest besti way
Makes it not say
.:, the⏎ found that -
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 9.37+ 790<|tool_call_end|><|assistant_end|><|tool_start|>1018<|tool_end|><|assistant_start|>The answer is 1
-
chat 9.2631 + 81<|tool_call_end|><|assistant_end|><|tool_start|>612<|tool_end|><|assistant_start|>The answer is 6
-
chat 8.9931 + 81<|tool_call_end|><|assistant_end|><|tool_start|>612<|tool_end|><|assistant_start|>The answer is 6
-
chat 8.99+ 201<|tool_call_end|><|assistant_end|><|tool_start|>1082<|tool_end|><|assistant_start|>The answer is 1
-
chat 8.95+ 947<|tool_call_end|><|assistant_end|><|tool_start|>1917<|tool_end|><|assistant_start|>The answer is 1
-
chat 8.8<|arg|>expression=6 + 4<|tool_call_end|><|assistant_end|><|tool_start|>10<|tool_end|><|assistant_start|>The answer is 1
-
story 8.79himself again, and he learned that being the best isn't always the most important thing.
-
chat 8.7756<|tool_end|><|assistant_start|>The answer is 1
-
chat 8.76expression=68 + 1<|tool_call_end|><|assistant_end|><|tool_start|>69<|tool_end|><|assistant_start|>The answer is 6
-
chat 8.6995 + 88<|tool_call_end|><|assistant_end|><|tool_start|>183<|tool_end|><|assistant_start|>The answer is 1
-
chat 8.667 + 750<|tool_call_end|><|assistant_end|><|tool_start|>967<|tool_end|><|assistant_start|>The answer is 9
-
chat 8.65+ 619<|tool_call_end|><|assistant_end|><|tool_start|>1280<|tool_end|><|assistant_start|>The answer is 1