I understand how packing is allowed in pretraining but I was looking for some clarification on how we are allowed to pack samples for SFT with ConstantLengthDataset. I see that an EOS token is put between samples https://github.com/huggingface/trl/blob/main/trl/trainer/utils.py#L567 but how are attention masks handled to make sure that the samples don't attend to each others context or is just putting EOS enough signal?

Could you also point me to the code for how multiple answers are handled in generation for loss calculations?