Feature #3290
Fires on “<|bos|>” (not labeled yet: read the examples)
- Fires on
- 0.45% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 35%
- chat 24%
- story 18%
- agent 17%
- python 6%
Makes the model say
stuck needed space girl+So took cool
Makes it not say
"```Oneutilssn3`
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 5.0848 + 739 = 10087<|bos|>Once upon a time
-
chat 4.88⏎8026 + 0 = 8026<|bos|><|system_start|>Answer with the
-
chat 4.80797 + 6 = 40803<|bos|><|system_start|>tools: calculator
-
agent 4.61⏎9881 + 5 = 9886<|bos|><|system_start|>tools: edit
-
chat 4.4725 + 2726 = 7051<|bos|><|system_start|>tools: calculator
-
agent 4.4792 + 384 = 89576<|bos|><|system_start|>tools: edit
-
chat 4.38⏎7719 + 3 = 7722<|bos|><|user_start|>What can you
-
worksheet 4.31279 and 21 is 3300.<|bos|>How much is
-
worksheet 4.2616⏎81 + 31 = 112<|bos|>The sum of
-
story 4.2653 + 5403 = 8056<|bos|>Once upon a time
-
story 4.1837 = 91540<|bos|>Tom and Mia were
-
agent 4.158 + 9838 = 54206<|bos|><|system_start|>tools: edit