Feature #4044
Fires on “⏎” (not labeled yet: read the examples)
- Fires on
- 0.14% of tokens
- Mostly in
- python
- Layer
- after block 4
Where it fires
- python 82%
- story 6%
- agent 4%
- worksheet 4%
- chat 4%
Makes the model say
<|bos|>ie withShe inside sun⏎ir
Makes it not say
_ repliedlab helpedse factsmallest helping
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
python 2.19max_item(nums):⏎ return max(nums)⏎```⏎<|bos|><|system_start|>Start every
-
python 2.14):⏎ return sum(nums) / len(nums)⏎```⏎⏎Bug: `
-
python 2.11):⏎ return sum(nums) / len(nums)⏎```⏎⏎Bug: `
-
python 2.1):⏎ return sum(nums) / len(nums)⏎```⏎<|bos|><|system_start|>tools:
-
python 2.1max_item(nums):⏎ return max(nums)⏎```⏎<|bos|># test_
-
python 2.08min_item(nums):⏎ return min(nums)⏎```⏎<|bos|><|user_start|>Could you
-
python 2.08⏎def largest(nums):⏎ return max(nums)⏎```⏎⏎Bug: `
-
python 2.07⏎def positive(x):⏎ return x > 0⏎```⏎⏎Bug: `
-
python 2.07next_number(x):⏎ return x + 1⏎```⏎⏎Bug: `
-
python 2.07):⏎ return sum(nums) / len(nums)⏎```⏎⏎Bug: `
-
python 2.07min_item(nums):⏎ return min(nums)⏎```⏎<|bos|><|user_start|>Mia has
-
python 2.06first_letter(s):⏎ return s[0]⏎```⏎⏎Bug: `