Feature #2376
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 0.66% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 50%
- python 33%
- story 10%
- chat 4%
- worksheet 3%
Makes the model say
woke sounds7 our thenitem equals)`.
Makes it not say
there_ sured special What favorite me
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
python 2.91⏎ return a⏎ return b⏎⏎⏎print(min_of(10, 9
-
chat 2.7is 152.<|assistant_end|><|bos|><|user_start|>Who won the last World Cup?
-
python 2.68nums):⏎ return min(nums)⏎⏎⏎print(smallest([1, 7,
-
agent 2.67won the World Cup?
-
python 2.66:⏎ return a⏎ return b⏎⏎⏎print(bigger(12, 9
-
chat 2.64bird.<|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>Who won the last World Cup?
-
python 2.63):⏎ return a * b⏎⏎⏎print(multiply(11, 9
-
chat 2.63with the triangle every day.<|assistant_end|><|bos|><|user_start|>Who won the last World Cup?
-
python 2.61import lower_of, difference⏎⏎print(lower_of(1, 3))
-
python 2.6(a + b) / 2⏎⏎⏎print(mean(10, 9
-
python 2.59):⏎ return a + b⏎⏎⏎print(sum_two(10, 1
-
chat 2.586⏎banana cloud⏎n⏎<|bos|><|user_start|>Who won the last World Cup?