Feature #3238
Fires on “=” (not labeled yet: read the examples)
- Fires on
- 0.51% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 50%
- chat 26%
- worksheet 11%
- story 10%
- python 3%
Makes the model say
<|tool_start|> enoughanged awayapeButbye<|think_start|>
Makes it not say
s S big-<|bos|> her Python Mr
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 2.14.⏎426+447: 6+7=13,
-
chat 2.12<|assistant_start|><|think_start|>096+667: 6+7=13,
-
chat 2.04<|assistant_start|><|think_start|>307+026: 7+6=13,
-
chat 2.04<|assistant_start|><|think_start|>307+026: 7+6=13,
-
worksheet 2.04=6, 6⏎7+6: 7+6=13,
-
chat 2.03<|assistant_start|><|think_start|>376+007: 6+7=13,
-
worksheet 2.036 = ?⏎07+26: 7+6=13,
-
worksheet 2.020007+7786: 7+6=13,
-
chat 2.01=6, 6⏎7+6: 7+6=13,
-
worksheet 1.986137+7866: 7+6=13,
-
chat 1.98<|assistant_start|><|think_start|>307+026: 7+6=13,
-
worksheet 1.94687+03536: 7+6=13,