Feature #990
Fires on “ be” (not labeled yet: read the examples)
- Fires on
- 0.06% of tokens
- Mostly in
- python
- Layer
- after block 4
Where it fires
- python 46%
- agent 34%
- story 16%
- chat 2%
- worksheet 1%
Makes the model say
treatwhere And<|tool_call_start|>defas explorecked
Makes it not say
, how( became need turned fast many
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 1.52beat the cream.
-
agent 1.52be `reverse`. It
-
story 1.52be doing a great job
-
agent 1.52be `positive`. It
-
story 1.52be fixed." The poor
-
story 1.52be clean, so he
-
story 1.26more careful when he went on his adventures, helping him to prevent further danger.
-
story 1.25knife and put it in her bag to keep everyone safe. Lily felt
-
story 1.23keeps beeping. Tom
-
story 1.2sign that said "Welcome!" There was a big party with all their friends. Tim was
-
chat 1.19his mom. Now, Tim knew he could be brave and help keep everyone safe.<|assistant_end|><|bos|>
-
story 1.17ious too. But then, both dogs became friends and played together with their toys. And they