Feature #247
Fires on “ add” (not labeled yet: read the examples)
- Fires on
- 0.26% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 72%
- agent 18%
- story 8%
- python 2%
- worksheet 0%
Makes the model say
l hand allt them numbers go everyone
Makes it not say
<|assistant_end|> upon`,commanddemo{zed`:
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 3.21can't help with that. I only know how to tell stories and add numbers.<|assistant_end|><|bos|>
-
chat 3.13can't help with that. I only know how to tell stories and add numbers.<|assistant_end|><|bos|>
-
chat 3.13can't help with that. I only know how to tell stories and add numbers.<|assistant_end|><|bos|>
-
chat 3.12can't help with that. I only know how to tell stories and add numbers.<|assistant_end|><|bos|>
-
chat 3.11can't help with that. I only know how to tell stories and add numbers.<|assistant_end|><|bos|>
-
chat 3.11can't help with that. I only know how to tell stories and add numbers.<|assistant_end|><|bos|>
-
chat 3.11can't help with that. I only know how to tell stories and add numbers.<|assistant_end|><|bos|>
-
chat 3.09can't help with that. I only know how to tell stories and add numbers.<|assistant_end|><|user_start|>
-
chat 3.09can't help with that. I only know how to tell stories and add numbers.<|assistant_end|><|user_start|>
-
chat 3.05can't help with that. I only know how to tell stories and add numbers.<|assistant_end|><|user_start|>
-
chat 3.04can't help with that. I only know how to tell stories and add numbers.<|assistant_end|><|bos|>
-
chat 3.03can't help with that. I only know how to tell stories and add numbers.<|assistant_end|><|user_start|>