Should the system message contribute to the SFT loss?

**Option B is correct.** During supervised fine-tuning, you want the model to learn one thing: how to produce the assistant turn given everything that came before it. The system message and the user message are inputs that condition the response; they are not things the model has to write at serving time. If you score the system and user spans, the model spends capacity learning to predict the prompt template and the user phrasing it has already been handed. That is wasted capacity. Worse, when many examples share the same system prompt, the model overfits to it and you see training loss drop suspiciously fast on early epochs. The fix is loss masking. In Hugging Face's convention you set the label ids for non-assistant tokens to negative one hundred, which tells the cross-entropy loss to skip them. The data collator builds these masked labels automatically when you wire up a chat template correctly. The net effect: only the assistant span's tokens contribute gradient. That is the signal you actually want the model to learn from.

**Answer B.** SFT teaches the model to generate the assistant response conditioned on the rest of the conversation. Scoring the system and user tokens would teach the model to also predict those, which is wasted capacity and a clean source of overfit. The mechanism is loss masking. The data collator builds a `labels` tensor parallel to `input_ids`, but for every token position that is not part of the assistant turn, the label is set to -100. PyTorch's cross-entropy loss treats -100 as 'ignore', so those positions contribute zero gradient. The forward pass still runs on the full sequence, the attention still attends to the conditioning tokens, but only assistant-span positions carry signal. Distractor A is the classic SFT bug. When you skip masking, training loss drops fast because the model trivially learns to copy the system prompt. The serving-time behaviour usually looks fine on short prompts but degrades on prompts that diverge from the training template. Distractor D conflates two unrelated layers. The chat template formats tokens; the labels tensor decides what gets scored. Llama 3, ChatML, Qwen, and Mistral all use the same masking convention.