Feature #1662
Fires on “<|assistant_end|>” (not labeled yet: read the examples)
- Fires on
- 0.59% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 56%
- chat 21%
- worksheet 13%
- story 5%
- python 5%
Makes the model say
` plus +Oapletterutfrom
Makes it not say
assert there moral feeling both now W few
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 3.31<|assistant_end|><|tool_start|>Edited number_
-
chat 3.31<|assistant_end|><|bos|>Lily loves to
-
chat 3.31<|assistant_end|><|user_start|>Add 1
-
agent 3.31<|assistant_end|><|tool_start|>Edited arith.
-
chat 3.31<|assistant_end|><|bos|>```python⏎
-
chat 3.31<|assistant_end|><|user_start|>Tom has
-
agent 3.31<|assistant_end|><|user_start|>What's the
-
agent 3.31<|assistant_end|><|tool_start|>Hello, world
-
agent 3.31<|assistant_end|><|tool_start|>Read file words
-
chat 3.31<|assistant_end|><|user_start|>Hi!<|user_end|>
-
chat 3.31<|assistant_end|><|user_start|>Plus 6
-
agent 3.31<|assistant_end|><|bos|><|system_start|>tools: