Feature #2890
Fires on “ ” (not labeled yet: read the examples)
- Fires on
- 0.68% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 66%
- chat 21%
- python 8%
- agent 4%
- story 2%
Makes the model say
to a herselfAnna other learned be cat
Makes it not say
idegememberaktestaei sure
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 3.4965+00: 5+0=5, 5547
-
worksheet 3.47+00400: 9+0=9, 9⏎39
-
python 3.41<|assistant_end|><|bos|>```python⏎def magnitude(x):⏎ if x < 0:⏎ return
-
chat 3.458+00: 8+0=8, 84⏎5
-
worksheet 3.39+03108: 0+8=8, 8⏎58
-
worksheet 3.33+00772: 5+2=7, 7⏎77
-
worksheet 3.2700+45: 0+5=5, 526⏎
-
worksheet 3.23+51801: 5+1=6, 6⏎00
-
agent 3.21<|arg|>content=def abs_value(x):⏎ if x < 0:⏎ return
-
chat 3.2112+00: 2+0=2, 29⏎1
-
worksheet 3.25+7561: 5+1=6, 6⏎37
-
python 3.14.py⏎def abs_value(x):⏎ if x < 0:⏎ return