Feature #534
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 0.10% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 54%
- chat 41%
- story 4%
- agent 1%
- python 1%
Makes the model say
What420Tim5irMom
Makes it not say
alsout⏎ nitude Forground missed "$
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 5.98⏎7+8: 7+8=15, 15814
-
chat 5.87+8: 8+8=16, 167<|think_end|>Sure
-
chat 5.81<|think_start|>8+9: 8+9=17, 17<|think_end|>The answer
-
worksheet 5.767+8: 7+8=15, 15791
-
worksheet 5.68⏎8+7: 8+7=15, 1592⏎
-
chat 5.66<|think_start|>9+9: 9+9=18, 18<|think_end|>The answer
-
worksheet 5.636+7: 6+7=13, 13810
-
worksheet 5.59+8: 4+8=12, 12⏎So
-
worksheet 5.59⏎7+7: 7+7=14, 14402
-
worksheet 5.58⏎7+8: 7+8=15, 15864
-
worksheet 5.55⏎9+8: 9+8=17, 17580
-
worksheet 5.53+8: 9+8+1=18, 18440