Feature #3510
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 0.16% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 36%
- story 30%
- chat 18%
- agent 9%
- python 7%
Makes the model say
<|tool_call_end|>>>> example here Come +Oncewords
Makes it not say
timeles wases awayLucy forgive would
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 5.9213⏎5+6: 5+6+1=12, 1
-
chat 5.7204⏎6+5: 6+5+1=12, 1
-
worksheet 5.71667+5825: 7+5=12, 2
-
chat 5.69<|user_end|><|assistant_start|><|think_start|>07+95: 7+5=12, 2
-
worksheet 5.67⏎007+395: 7+5=12, 2
-
worksheet 5.63597+0025: 7+5=12, 2
-
worksheet 5.4907+99835: 7+5=12, 2
-
chat 5.4848⏎5+7: 5+7=12, 1
-
worksheet 5.396+0075: 6+5+1=12, 2
-
worksheet 5.395+6496: 5+6+1=12, 2
-
worksheet 5.39⏎885+687: 5+7=12, 2
-
worksheet 5.3355+69667: 5+7=12, 2