Feature #1620
Fires on “ =” (not labeled yet: read the examples)
- Fires on
- 0.45% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 36%
- worksheet 23%
- python 19%
- story 16%
- chat 7%
Makes the model say
+[ doubleild [ave exclamation return
Makes it not say
<|assistant_end|> tinyBIt then It called time
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 3.7843892 shells.⏎6 + 26 = 32⏎
-
worksheet 3.73get 84252.⏎7 + 26 = ?⏎07
-
worksheet 3.59get 95184.⏎6 + 82 = 88⏎
-
worksheet 3.52has 758 apples.⏎95 + 92 = ?⏎95
-
worksheet 3.48get 46277.⏎86 + 8 = 94⏎
-
worksheet 3.45121949.⏎6 + 810 = 816
-
worksheet 3.447 = 7771.⏎5 + 78 = 83⏎
-
worksheet 3.4460518.⏎45 + 32 = 77⏎
-
worksheet 3.42is 5366.⏎47 + 25 = 72<|bos|>
-
chat 3.37ful dog named Spot.<|assistant_end|><|bos|><|user_start|>77 + 30 = ?<|user_end|><|assistant_start|>The
-
worksheet 3.35has 170 flowers.⏎88 + 18 = ?⏎88
-
worksheet 3.34= 15517.⏎2 + 68 = 70<|bos|>