Feature #1857
Fires on “ prints” (not labeled yet: read the examples)
- Fires on
- 0.31% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 89%
- story 6%
- chat 2%
- worksheet 2%
- python 1%
Makes the model say
function trainedcontent<|user_end|><|assistant_end|>ake passesappy
Makes it not say
how the tree st he she rock
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 3.88py`: `incldes` should be `includes`. It now prints:⏎```⏎
-
agent 3.84py`: `douubled` should be `doubled`. It now prints:⏎```⏎
-
agent 3.81has__item` should be `has_item`. It now prints:⏎```⏎
-
agent 3.78proudct` should be `product`. It now prints:⏎```⏎
-
agent 3.76`: `whsper` should be `whisper`. It now prints:⏎```⏎
-
agent 3.75.py`: `sout` should be `shout`. It now prints:⏎```⏎
-
agent 3.75_eveen` should be `is_even`. It now prints:⏎```⏎
-
agent 3.75py`: `repet` should be `repeat`. It now prints:⏎```⏎
-
agent 3.73py`: `shhout` should be `shout`. It now prints:⏎```⏎
-
agent 3.7py`: `evn` should be `even`. It now prints:⏎```⏎
-
agent 3.7a_sitem` should be `has_item`. It now prints:⏎```⏎
-
agent 3.7py`: `filp` should be `flip`. It now prints:⏎```⏎