Feature #2555
Fires on “ max” (not labeled yet: read the examples)
- Fires on
- 1.30% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 71%
- python 24%
- story 3%
- chat 1%
- worksheet 1%
Makes the model say
([_ positiveab r seaOhe
Makes it not say
:// sheanceonewayAdded])Rename
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 3.13import is_odd, raise_to⏎⏎ Line 5: assert raise_to(1
-
agent 2.92, raise_to, sum_to⏎⏎ Line 4: assert raise_to(5
-
agent 2.861: from textlib import final_letter⏎⏎ Line 3: assert final_letter("lemon
-
agent 2.79import max_of, next_number⏎⏎ Line 3: assert max_of(1
-
agent 2.69import double_all, max_item⏎⏎ Line 5: assert max_item([3
-
agent 2.671: from textlib import num_chars⏎⏎ Line 3: assert num_chars("pizza
-
agent 2.66, add_all, mean_list⏎⏎ Line 3: assert count([7,
-
agent 2.63from items import has_item, first⏎⏎ Line 3: assert has_item([9
-
agent 2.63import first, includes, max_item⏎⏎ Line 5: assert max_item([6
-
agent 2.531: from arith import positive, minus⏎⏎ Line 3: assert positive(-1) ==
-
agent 2.521: from listlib import max_item⏎⏎ Line 3: assert max_item([8
-
agent 2.511: from helpers import smaller, count⏎⏎ Line 5: assert count([6,