Feature #3208
Fires on “5” (not labeled yet: read the examples)
- Fires on
- 0.59% of tokens
- Mostly in
- python
- Layer
- after block 4
Where it fires
- python 38%
- agent 25%
- worksheet 22%
- chat 13%
- story 4%
Makes the model say
up numbers passes PythoneRenameGood<|user_start|>
Makes it not say
saw why foundoWhy wind big play
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 4.3156 and 44
-
worksheet 4.315657 plus
-
worksheet 4.31560.⏎What
-
chat 4.3156<|tool_end|><|assistant_start|>The answer
-
worksheet 4.3156: 9+
-
worksheet 4.1755246 plus
-
worksheet 4.175534 +
-
worksheet 4.175, 56
-
worksheet 4.175.⏎What is
-
agent 4.175: def lower_
-
worksheet 4.17547=⏎
-
worksheet 4.175: 8+