Feature #2177
Fires on “ it” (not labeled yet: read the examples)
- Fires on
- 0.56% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 47%
- story 34%
- chat 18%
- python 2%
- worksheet 1%
Makes the model say
.:self know, ` andOnce
Makes it not say
pathrite Python=**/*. playing<|tool_call_start|>ence)`.
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 3.05answer is 108.<|assistant_end|><|user_start|>Add 61 to it.<|user_end|><|assistant_start|><|tool_call_start|>
-
chat 3.04answer is 116.<|assistant_end|><|user_start|>Add 56 to it.<|user_end|><|assistant_start|>The
-
chat 2.99answer is 149.<|assistant_end|><|user_start|>Add 18 to it.<|user_end|><|assistant_start|>The
-
chat 2.98The answer is 78.<|assistant_end|><|user_start|>Add 56 to it.<|user_end|><|assistant_start|><|think_start|>
-
chat 2.96answer is 183.<|assistant_end|><|user_start|>Add 56 to it.<|user_end|><|assistant_start|><|tool_call_start|>
-
chat 2.96<|assistant_end|><|user_start|>Add 1 to it.<|user_end|><|assistant_start|><|think_start|>
-
story 2.94saw a big jar. It was deep and had a funny picture on it. Tom's mom
-
story 2.94ater is boring. It has a plain star on it. Mine is
-
chat 2.89.<|assistant_end|><|user_start|>Add 23 to it.<|user_end|><|assistant_start|>The
-
chat 2.89<|assistant_start|>The answer is 6.<|assistant_end|><|user_start|>Add 53 to it.<|user_end|><|assistant_start|><|think_start|>
-
chat 2.88The answer is 24.<|assistant_end|><|user_start|>Add 80 to it.<|user_end|><|assistant_start|><|think_start|>
-
chat 2.87<|assistant_start|>The answer is 7.<|assistant_end|><|user_start|>Add 50 to it.<|user_end|><|assistant_start|>The