u can do the following :

1. Masking + Loss Adjustment:

   • Pad targets with zeros (or any value),

   • Use a mask during training to ignore padded positions in the loss calculation.

2. Avoid Ambiguous Padding:

   • Instead of 0.0 (which may be valid), use a clearly invalid float like -9999.0,

   • Then apply masking to ignore this during training and evaluation.

3. Use Sequence Models:

   • Models like RNNs/Transformers naturally handle variable-length outputs, especially with attention and masking.

4. Dynamic Output Generation:

   • Consider seq2seq regression with an end-of-sequence (EOS) token (predict a flag alongside each value indicating end).

The key: combine padding with masking, and never penalize the model for predictions on padded areas.