Feature #1281
Fires on “ ” (not labeled yet: read the examples)
- Fires on
- 3.37% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 38%
- worksheet 28%
- chat 22%
- python 11%
- story 1%
Makes the model say
102 atOnce (Yes
Makes it not say
, Lily one walked there up way as
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 6.23played with the toy all day. The other dog saw this and felt envious too.
-
chat 6.08<|user_start|>Tell me a short story.<|user_end|><|assistant_start|>Bob was feeling excited. He was going to
-
chat 5.72in the park. He was jumping from
-
story 5.7big dog named Max. Max was very fast and strong. Tom felt envious of Max
-
story 5.66on the stay. Flip was very worried and regretful. Quickly
-
chat 5.65was a grumpy boat. The boat was always
-
story 5.63his friend, Sue, eating a big ice cream cone. Tim felt envious. He
-
story 5.53lunch. Whenever he did this, Jack would be so envious that he
-
chat 5.36a tree. He looked at his
-
python 5.27python⏎def sum_up_to(n):⏎ acc = 0⏎ for i
-
python 5.18⏎⏎⏎def count_vowels(s):⏎ k = 0⏎ for c
-
story 5.17, the dog saw another dog with a big toy. The dog felt envious. He