Feature #1308
Fires on “ sum” (not labeled yet: read the examples)
- Fires on
- 0.76% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 66%
- worksheet 12%
- python 8%
- chat 8%
- story 6%
Makes the model say
_merdit % awaylessase
Makes it not say
toin help item one sat cloud helped
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 4.58You are mini, a small and friendly assistant.<|system_end|><|user_start|>Find the sum of 75
-
chat 4.57up on its own.<|bos|><|user_start|>Find the sum of 90
-
chat 4.53do.<|bos|><|system_start|>You are a helpful assistant.<|system_end|><|user_start|>Find the sum of 0 and
-
chat 4.5abilities<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>Find the sum of 82
-
chat 4.49<|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>Find the sum of 41
-
chat 4.47. And the ant never complained again.<|bos|><|user_start|>Find the sum of 63
-
chat 4.44they were now friends.<|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>Find the sum of 6 and
-
chat 4.38: calculator<|system_end|><|system_start|>Do not use the calculator.<|system_end|><|user_start|>Find the sum of 22
-
worksheet 4.37The fat baby nodded, and hugged her toy tight.<|bos|>Find the sum of 62
-
chat 4.36: calculator<|system_end|><|system_start|>Do not use the calculator.<|system_end|><|user_start|>Find the sum of 68
-
chat 4.35<|system_start|>You are a helpful assistant.<|system_end|><|user_start|>Find the sum of 93
-
chat 4.34can only tell short stories and add numbers.<|assistant_end|><|bos|><|user_start|>Find the sum of 9 and