Probabilistic One Hot Encoding

Hello !

In my multiclass dataset, I have duplicates, but with different labels.
For exemple, I could have the three samples: “A” → 1, “A” → 1, “A” → 2.
I was wondering if I could merge them in one sample ? Instead of having two one hot encoding softmax [1,0] and [0,1], I’d like to have just one [2/3, 1/3].
Is it possible with Transformers library ?

Thank you in advance !