Feature #624
Fires on “ ” (not labeled yet: read the examples)
- Fires on
- 0.85% of tokens
- Mostly in
- python
- Layer
- after block 4
Where it fires
- python 47%
- agent 45%
- chat 5%
- worksheet 2%
- story 2%
Makes the model say
,ine for first you your I maybe
Makes it not say
waslistlib upon<|bos|> happened happen],<|assistant_end|>
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 3.27, a coding agent harness.<|system_end|><|user_start|>Set SCORES to [2, 3, 3
-
python 3.23, 8, 9]⏎print(head([6, 2, 0
-
agent 3.28, 9]<|arg|>newString=VALUES = [3, 7, 4
-
python 3.16import tail, min_item⏎⏎print(tail([6, 2, 7
-
python 3.13nums)⏎⏎⏎SCORES = [2, 2,
-
agent 3.133, 5]<|arg|>newString=SCORES = [4, 5, 5
-
agent 3.1=print(tial([9, 0, 4
-
python 3.04tail⏎⏎assert tail([8, 2,
-
agent 3.01, a coding agent harness.<|system_end|><|user_start|>Set VALUES to [8, 2, 3
-
python 3.0, 2, 4]))⏎print(evens([1, 5, 3
-
python 2.99is_odd⏎⏎assert last([8, 8,
-
agent 2.96<|arg|>newString=print(max_item([5, 2, 2