Feature #3463
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 0.56% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 63%
- chat 16%
- agent 13%
- story 7%
- python 1%
Makes the model say
+ square decided took I whoat dec
Makes it not say
lived<|bos|>(- `ed their``` never
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 2.65, 58082⏎So 6245 + 51
-
worksheet 2.485, 57077⏎So 1578 + 55
-
worksheet 2.46=3, 37666⏎So 843 + 36
-
worksheet 2.459, 96590⏎So 2951 + 93
-
worksheet 2.44, 102710⏎So 5017 + 97
-
worksheet 2.433, 36665⏎So 5155 + 31
-
worksheet 2.38=6, 67058⏎So 869 + 66
-
worksheet 2.375, 58892⏎So 2781 + 56
-
story 2.36answer is 6.<|assistant_end|><|bos|>Once upon a time there were two friends, a rabbit and
-
worksheet 2.353, 33515⏎So 4961 + 28
-
worksheet 2.35, 24138⏎So 23846 + 29
-
story 2.35ning a big meal.<|bos|>Once upon a time, there were two friends, John and Lisa