Feature #2019
Fires on “ evens” (not labeled yet: read the examples)
- Fires on
- 0.33% of tokens
- Mostly in
- python
- Layer
- after block 4
Where it fires
- python 86%
- agent 9%
- story 4%
- worksheet 0%
- chat 0%
Makes the model say
, small sherreed and met
Makes it not say
contentscriptRenameddemo<|user_end|>ananaICENSE prelude
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
python 7.74⏎⏎# test_listlib.py⏎from listlib import includes, first, average_list
-
python 7.65⏎⏎# test_stats.py⏎from stats import contains, biggest⏎⏎assert contains
-
python 7.65⏎⏎# test_listlib.py⏎from listlib import head, evens⏎⏎assert head
-
python 7.64# test_stats.py⏎from stats import has_item, sum_list⏎⏎
-
python 7.6⏎⏎# test_lists.py⏎from lists import last, evens⏎⏎assert last
-
python 7.41.<|bos|># test_listlib.py⏎from listlib import total, min_item⏎⏎
-
python 7.4⏎⏎# test_listlib.py⏎from listlib import first, smallest⏎⏎assert first
-
python 7.25⏎⏎# test_words.py⏎from words import quiet, exclaim⏎⏎assert quiet
-
python 7.22# test_lists.py⏎from lists import max_item, double_all⏎⏎
-
python 7.22# test_items.py⏎from items import keep_even, head⏎⏎assert keep
-
python 7.15⏎⏎# test_text.py⏎from text import whisper, welcome⏎⏎assert whisper
-
python 7.14<|assistant_end|><|bos|># test_words.py⏎from words import exclaim, echo, last_letter