Feature #898
Fires on “⏎” (not labeled yet: read the examples)
- Fires on
- 0.68% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 73%
- chat 22%
- story 3%
- python 2%
- worksheet 1%
Makes the model say
closer okay_ should<|system_end|>Trueanceap
Makes it not say
the so with very thangreetings run balls
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.95s, now I can have my delicious mouse!"<|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>Add
-
agent 2.91) / len(nums)⏎3: ⏎4: ⏎5: def sum
-
agent 2.88return x * x * x⏎3: ⏎4: ⏎5: def backwards
-
chat 2.88can only tell short stories and add numbers.<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|system_start|>Do not
-
agent 2.86return s[::1]⏎3: ⏎4: ⏎5: def hello
-
agent 2.86a % b == 0⏎3: ⏎4: ⏎5: def raise
-
chat 2.86later, we can write with a real pen."<|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>Mia has
-
agent 2.85return min(nums)⏎3: ⏎4: ⏎5: def hello
-
agent 2.85return max(nums)⏎3: ⏎4: ⏎5: def evens
-
agent 2.85) / len(nums)⏎3: ⏎4: ⏎5: def how
-
agent 2.85return a + "" + b⏎3: ⏎4: ⏎5: def backwards
-
agent 2.84a % b == 0⏎3: ⏎4: ⏎5: def includes