Feature #1914
Fires on “ated” (not labeled yet: read the examples)
- Fires on
- 0.22% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 27%
- worksheet 23%
- python 18%
- story 17%
- chat 15%
Makes the model say
k here but fish I believe But chair
Makes it not say
readedtxt groundone again goodbye other
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 1.73a and b separated by a space.
-
python 1.5: `join_words` should return a and b separated by a space,
-
agent 1.49<|assistant_start|>`combine(a, b)` returns a and b separated by a space.
-
python 1.45.<|assistant_end|><|bos|>Bug: `combine` should return a and b separated by a space,
-
agent 1.41b):⏎6: return a + " " + b⏎7
-
story 1.41and saw some water. He thought, "I can pour the water on the flowers!"
-
python 1.39def join_words(a, b):⏎ return a + " " + b⏎```
-
python 1.39def join_words(a, b):⏎ return a + " " + b⏎```
-
python 1.38words(a, b):⏎ return b + " " + a⏎
-
chat 1.33be careful." Buddy barked and nodded his head. He wanted to keep Sue safe.⏎Together
-
chat 1.32Tim and the tree became best friends. They spent many days playing and giving hugs to each
-
chat 1.31, 02⏎7+2: 7+2+1=10,