Feature #2171
Fires on “ ” (not labeled yet: read the examples)
- Fires on
- 3.98% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 54%
- chat 20%
- agent 15%
- python 8%
- story 3%
Makes the model say
Fix eched1 stillLookreadBut
Makes it not say
sunny alllast# Thestringsow do
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 8.3sum of 3970 and 9812 is 1378
-
worksheet 8.18+ 5 = 13⏎99 plus 9 equals 108.
-
worksheet 7.87<|bos|>89651 plus 69893 equals 1595
-
worksheet 7.8684.⏎894 + 4976 = 5870
-
worksheet 7.85.⏎485 plus 7983 equals 8468
-
worksheet 7.84⏎99345 plus 28969 equals 1283
-
worksheet 7.83of 2997 and 8947 is 11944
-
worksheet 7.83add 5955 and 955, you get 6910
-
worksheet 7.6169032 and 9969, you get 7900
-
worksheet 7.57sum of 4935 and 6998 is 1193
-
worksheet 7.44test_stats.py`.<|assistant_end|><|bos|>79 plus 9 is 88.⏎
-
worksheet 7.3448761 and 59709 is 1084