Feature #2099
Fires on “e” (not labeled yet: read the examples)
- Fires on
- 0.71% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 50%
- python 41%
- story 5%
- chat 3%
- worksheet 1%
Makes the model say
saidert saylllpg com next
Makes it not say
<|assistant_end|> ` config{<|tool_start|>list example
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
python 7.81reverses s. For example, `reverse("apple")` returns `"elppa"`.⏎
-
python 7.7s reversed. For example, `flip("python")` returns `"nohtyp
-
python 7.67return s[::-1]⏎⏎$ python3 main.py⏎eert⏎$ python
-
python 7.51") == "azzip"⏎assert flip("apple") == "elppa"⏎
-
python 7.47s reversed. For example, `flip("tree")` returns `"eert"`.⏎⏎
-
python 7.36(backwards("hello"))⏎⏎$ python3 main.py⏎olleh⏎
-
python 7.22("star") == 1⏎assert flip("tree") == "eert"⏎assert
-
python 6.95to run the tests.⏎⏎$ python3 demo.py⏎edoc⏎$ python
-
python 6.76return s[-1]⏎⏎$ python3 demo.py⏎esuoh⏎o
-
python 6.57"pizzapizza"⏎assert flip("code") == "edoc"⏎assert
-
python 6.5") == "ananab"⏎assert backwards("robot") == "tobor"⏎
-
python 6.49== "Hello, Max!"⏎assert reverse("moon") == "noom"⏎assert