Feature #1649
Fires on “⏎” (not labeled yet: read the examples)
- Fires on
- 0.80% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 54%
- story 18%
- worksheet 17%
- python 7%
- chat 4%
Makes the model say
+ == import<|tool_call_start|><|user_end|>ve who for
Makes it not say
( Then<|bos|>ep by largest biggest friends
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.32⏎4: ⏎5: def upper_case(s):⏎6: return
-
agent 2.284: ⏎5: def loud(s):⏎6: return
-
agent 2.031-6⏎1: def size(s):⏎2: return
-
agent 2.02⏎8: ⏎9: def halve(x):⏎10:
-
agent 2.02⏎5: def vowels(s):⏎6: k
-
agent 2.02: ⏎4: ⏎5: def exclaim(s):⏎6: return
-
agent 2.0⏎1: def num_chars(s):⏎2: return
-
story 1.94saw the beautiful universe together. They knew they would be best friends forever
-
agent 1.93: ⏎4: ⏎5: def initial(s):⏎6: return
-
agent 1.93: def shout(s):⏎<|tool_end|><|assistant_start|><|tool_call_start|>read
-
story 1.92thanks to her soft friend Lily. Sue hugged Lily and knew they would always take care of
-
agent 1.91largest(nums):⏎10: