Feature #3989
Fires on “ ” (not labeled yet: read the examples)
- Fires on
- 0.42% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 38%
- chat 21%
- story 19%
- agent 13%
- python 9%
Makes the model say
- ): betterTim was + very
Makes it not say
<|user_start|>$t<|assistant_end|> would togetherlower
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 3.338⏎5+7: 5+7=12, 1248
-
worksheet 3.33⏎8+7: 8+7=15, 1592
-
worksheet 3.264+7: 4+7+1=12, 1203
-
worksheet 3.262⏎7+7: 7+7=14, 1440
-
worksheet 3.224⏎7+8: 7+8=15, 1581
-
chat 3.224+7: 4+7+1=12, 121<|think_end|>
-
worksheet 3.2?⏎7+4: 7+4=11, 11⏎So
-
chat 3.19<|assistant_start|><|think_start|>8+9: 8+9=17, 17<|think_end|>The
-
worksheet 3.176+7: 6+7=13, 1381
-
worksheet 3.174+7: 4+7+1=12, 1247
-
chat 3.15+8: 8+8=16, 167<|think_end|>
-
chat 3.158⏎4+7: 4+7=11, 1138