Feature #3769
Fires on “5” (not labeled yet: read the examples)
- Fires on
- 0.88% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 90%
- worksheet 5%
- story 2%
- chat 2%
- python 1%
Makes the model say
assert_tiger<|tool_call_end|>onfig Python>>>+
Makes it not say
⏎it, In way into!!"
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 3.797: ⏎8: ⏎9
-
agent 3.76min(nums)⏎3: ⏎4: ⏎5: def last(
-
agent 3.75nums]⏎3: ⏎4: ⏎5: def average_
-
agent 3.73in nums]⏎3: ⏎4: ⏎5: def keep_
-
agent 3.73: ⏎4: ⏎5: def exclaim(
-
agent 3.73: ⏎4: ⏎5: def factorial(
-
agent 3.66: ⏎4: ⏎5: def join_
-
agent 3.66: ⏎4: ⏎5: def smallest(
-
agent 3.65+ name + "!"⏎7: ⏎8: ⏎9: def initial(
-
agent 3.634: ⏎5: def loud(
-
agent 3.638: ⏎9: def raise_
-
agent 3.638: ⏎9: WORD = "