Feature #1625
Fires on “ b” (not labeled yet: read the examples)
- Fires on
- 0.55% of tokens
- Mostly in
- python
- Layer
- after block 4
Where it fires
- python 63%
- agent 34%
- story 2%
- chat 1%
- worksheet 0%
Makes the model say
help⏎⏎ are be l thank large has
Makes it not say
po typoCreatedusedowed "" comingwords
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
python 4.23in items⏎⏎⏎def biggest(nums):⏎ return max(nums)⏎⏎# test
-
python 4.22⏎⏎⏎def max_item(nums):⏎ return max(nums)⏎⏎⏎def combine
-
python 4.21_words(a, b):⏎ return a + " " + b⏎⏎# textlib
-
python 4.2_words(a, b):⏎ return a + " " + b⏎```⏎⏎
-
python 4.2_words(a, b):⏎ return a + " " + b⏎```⏎⏎
-
python 4.18acc⏎⏎⏎def multiply(a, b):⏎ return a * b⏎⏎$ python
-
agent 4.15list(nums):⏎ return sum(nums) / len(nums)<|tool_call_end|><|assistant_end|><|tool_start|>Edited
-
python 4.13b⏎⏎⏎def times(a, b):⏎ return a * b⏎⏎# words
-
agent 4.11list(nums):⏎ return sum(nums) / len(nums)⏎<|tool_call_end|><|assistant_end|><|tool_start|>
-
python 4.1b⏎⏎⏎def product(a, b):⏎ return a * b⏎⏎# utils
-
python 4.06, b):⏎ if a > b:⏎ return a⏎ return b⏎```⏎<|bos|>
-
python 4.040]⏎⏎⏎def biggest(nums):⏎ return max(nums)⏎⏎⏎print(