Feature #368
Fires on “ for” (not labeled yet: read the examples)
- Fires on
- 0.63% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 55%
- agent 28%
- python 7%
- story 6%
- worksheet 3%
Makes the model say
assertant tests sl > perfect))ared
Makes it not say
6 of417 have83
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.93.<|bos|><|user_start|>Can you add 38 and 327 for me?<|user_end|><|assistant_start|>
-
chat 2.91: calculator<|system_end|><|user_start|>Can you add 51 and 30 for me?<|user_end|><|assistant_start|>
-
chat 2.870.<|assistant_end|><|bos|><|user_start|>Can you add 1 and 93 for me?<|user_end|><|assistant_start|>
-
chat 2.83you today?<|assistant_end|><|user_start|>Can you add 5 and 45 for me?<|user_end|><|assistant_start|>
-
chat 2.8321301.<|bos|><|user_start|>8+5=<|user_end|><|assistant_start|>The answer is 1
-
chat 2.79<|system_end|><|user_start|>Can you add 798 and 551 for me?<|user_end|><|assistant_start|>
-
chat 2.78for another story soon.<|assistant_end|><|user_start|>Play some music for me.<|user_end|><|assistant_start|>
-
chat 2.78friendly assistant.<|system_end|><|user_start|>Can you add 7 and 47 for me?<|user_end|><|assistant_start|>
-
chat 2.78<|assistant_end|><|bos|><|user_start|>Can you add 32 and 340 for me?<|user_end|><|assistant_start|>
-
chat 2.77_utils.py`.<|assistant_end|><|bos|><|user_start|>0 + 4<|user_end|><|assistant_start|>The answer is 4
-
chat 2.73<|bos|><|system_start|>Answer with the number only.<|system_end|>
-
chat 2.73<|bos|><|system_start|>Answer with the number only.<|system_end|>