Feature #2475
Fires on “ things” (not labeled yet: read the examples)
- Fires on
- 0.12% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 45%
- story 23%
- python 20%
- worksheet 9%
- chat 4%
Makes the model say
+backwards://<|tool_start|>[::-<|system_start|>:` unexpected
Makes it not say
py about toious down's of kids
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 1.57box. Only a note that said: "You should not take things that are not yours
-
story 1.48Flip learnt his lesson: that it's not okay to take things without our parents c
-
story 1.47give back the food. The car learned that it was bad to take things without paying.
-
story 1.46learned a hard lesson today," she said. "You should not take things that are not yours
-
story 1.43time. Sally had learned an important lesson - sometimes you need to take things slow and make sure
-
story 1.39to her heart, knowing that it was something special that she had earned.<|bos|><|user_start|>
-
story 1.36. Ben and Tom were happy and safe. They learned that sometimes, something scary can help you
-
story 1.31of the story is that it is important to be careful and never take a wrong turn, other
-
story 1.31. Sometimes pepper can be smelly, but it will make our soup taste good."⏎They
-
story 1.3Tom learned that sometimes, even when things seem bad, they
-
story 1.29. Doing your work is important and being lazy won't take you very far in life.
-
story 1.28is that sharing can make everyone happy, even when things don't go as planned.<|bos|>