Feature #2294
Fires on “ return” (not labeled yet: read the examples)
- Fires on
- 1.60% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 94%
- python 3%
- chat 1%
- story 1%
- worksheet 1%
Makes the model say
[-()`<|think_end|> **])`'? length
Makes it not say
of2ies. tree they way house
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 3.111: def sum_to(n):⏎2: acc = 0⏎3
-
agent 3.021: def sum_to(n):⏎2: acc = 0⏎3
-
agent 2.97-2⏎1: def size(s):⏎2: return len(s)
-
agent 2.9510⏎1: def evens(nums):⏎2: return [n for n
-
agent 2.94⏎1: def sub_one(x):⏎2: return x - 1
-
agent 2.93def sum_up_to(n):⏎2: acc = 0⏎3
-
agent 2.921: def sum_to(n):⏎2: acc = 0⏎3
-
agent 2.9114⏎1: def cubed(x):⏎2: return x * 3
-
agent 2.9⏎1: def final_letter(s):⏎2: return s[-1]
-
agent 2.89def sum_up_to(n):⏎2: acc = 0⏎3
-
agent 2.88⏎1: def sub_one(x):⏎2: return x + 1
-
agent 2.86-2⏎1: def square(x):⏎2: return x * 2