Feature #1626
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 0.36% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 48%
- chat 25%
- agent 24%
- story 2%
- python 1%
Makes the model say
0<|tool_call_end|> should1 so taking three take
Makes it not say
lidRenameu://magLz
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 9.56⏎5+5: 5+5=10, 1
-
chat 9.375: 5+5=10, 1
-
worksheet 9.21965+4685: 5+5=10, 0
-
worksheet 9.139286⏎5+5: 5+5=10, 1
-
chat 9.11<|user_end|><|assistant_start|><|think_start|>05+45: 5+5=10, 0
-
worksheet 8.99801⏎5+5: 5+5=10, 1
-
worksheet 8.9965+00585: 5+5=10, 0
-
worksheet 8.82⏎035+515: 5+5=10, 0
-
worksheet 8.81755+5295: 5+5=10, 0
-
chat 8.814⏎05+45: 5+5=10, 0
-
worksheet 8.72869⏎5+5: 5+5=10, 1
-
chat 8.64<|think_start|>055+715: 5+5=10, 0