Feature #3883
Fires on “⏎” (not labeled yet: read the examples)
- Fires on
- 1.29% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 73%
- python 21%
- story 4%
- worksheet 2%
- chat 1%
Makes the model say
```⏎ moral<|tool_call_start|>ed example endved
Makes it not say
am around named off out)` twice
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.77: ⏎9: def factorial(n):⏎10:
-
agent 2.76: ⏎4: ⏎5: def smallest(nums):⏎6: return
-
agent 2.72⏎4: ⏎5: def sum_list(nums):⏎6: result
-
agent 2.634: ⏎5: def loud(s):⏎6: return
-
python 2.58- b⏎⏎# script.
-
agent 2.57: ⏎4: ⏎5: def factorial(n):⏎6: result
-
agent 2.52: ⏎4: ⏎5: def last(items):⏎6: return
-
agent 2.52: ⏎4: ⏎5: def exclaim(s):⏎6: return
-
agent 2.51: ⏎6: ⏎7: def smallest(nums):⏎8: return
-
agent 2.5⏎8: ⏎9: def halve(x):⏎10:
-
agent 2.49⏎11: ⏎12: def smallest(nums):⏎13:
-
agent 2.47⏎4: ⏎5: def upper_case(s):⏎6: return