Feature #1029
Fires on “ has” (not labeled yet: read the examples)
- Fires on
- 0.80% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 50%
- chat 36%
- python 10%
- story 2%
- agent 2%
Makes the model say
numbers_)newString<|assistant_end|> every name
Makes it not say
>>> donNo doing youaringlar
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
python 1.99f` should return half of x, rounded down, but it has `return x //
-
python 1.9` should add up the numbers from 0 to n, but it has `for i in
-
chat 1.89+0=3, 3⏎4+0: 4+0
-
worksheet 1.88+0+1=8, 83⏎6+0: 6+0
-
python 1.84f` should return half of x, rounded down, but it has `return x *
-
python 1.83f` should return half of x, rounded down, but it has `return x //
-
python 1.82: `last` should return the last element of items, but it has `return items[
-
python 1.81: `last` should return the last element of items, but it has `return items[
-
worksheet 1.8+1=13, 300⏎6+0: 6+0
-
python 1.79.<|assistant_end|><|bos|>Bug: `flip` should reverse s, but it has `return s[
-
worksheet 1.796+1=12, 23⏎9+0: 9+0
-
python 1.78of items, but it has `return items[