Feature #3926
Fires on โ1โ (not labeled yet: read the examples)
- Fires on
- 0.66% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 39%
- agent 33%
- chat 15%
- story 10%
- python 4%
Makes the model say
<|bos|> Bts show.nd<|assistant_end|>
Makes it not say
ocked returns)) == but?"+ gets
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 3.11: 4+2+1=7,
-
worksheet 3.083: 7+3+1=11,
-
worksheet 3.062: 4+2+1=7,
-
story 2.99repeat what the clock said
-
worksheet 2.9: 8+2+1=11,
-
worksheet 2.9347: 3+7+1=11,
-
worksheet 2.89: 5+7+1=13,
-
worksheet 2.872+93: 2+3+1=6,
-
worksheet 2.82: 0+6+1=7,
-
worksheet 2.8: 4+3+1=8,
-
worksheet 2.798: 4+8+1=13,
-
worksheet 2.79, 0โ0+3: 0+3+1=4,