Feature #3899
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 0.26% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 54%
- chat 24%
- story 11%
- python 6%
- agent 4%
Makes the model say
3⏎ maybe:`words instead:`. again
Makes it not say
nameds aant was_he One
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 7.53<|user_end|><|assistant_start|><|think_start|>16+77: 6+7=13, 3
-
chat 7.46<|think_start|>376+007: 6+7=13, 3
-
chat 7.39<|think_start|>096+667: 6+7=13, 3
-
chat 7.31846+767: 6+7=13, 3
-
worksheet 7.3007+7786: 7+6=13, 3
-
worksheet 7.216+7: 6+7=13, 1
-
chat 7.139, 9⏎6+7: 6+7=13, 1
-
worksheet 7.1⏎626+007: 6+7=13, 3
-
chat 7.096, 6⏎7+6: 7+6=13, 1
-
worksheet 7.07716+0097: 6+7=13, 3
-
chat 7.0717: 6+7=13, 3
-
chat 7.04<|think_start|>446+047: 6+7=13, 3