Feature #2391
Fires on “=” (not labeled yet: read the examples)
- Fires on
- 0.91% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 32%
- worksheet 31%
- agent 26%
- python 6%
- story 5%
Makes the model say
everureatedpositiveappydge(
Makes it not say
The "01 the liked good was
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 2.577⏎000+600: 0+0=0, 0
-
worksheet 2.573⏎000+570: 0+0=0, 0
-
chat 2.55for?<|user_end|><|assistant_start|>Sorry, I'm a tiny model: I can only tell short stories and
-
worksheet 2.57⏎000+100: 0+0=0, 0
-
chat 2.48, 8⏎00+50: 0+0=0, 0
-
worksheet 2.480000+1080: 0+0=0, 0
-
chat 2.43morning.<|user_end|><|assistant_start|>Sorry, I'm a tiny model: I can only tell short stories and
-
worksheet 2.420000+3000: 0+0=0, 0
-
chat 2.42for?<|user_end|><|assistant_start|>Sorry, I'm a tiny model: I can only tell short stories and
-
worksheet 2.41529⏎00+80: 0+0=0, 0
-
chat 2.41up?<|user_end|><|assistant_start|>Sorry, I'm a tiny model: I can only tell short stories and
-
worksheet 2.411⏎000+820: 0+0=0, 0