I have a question regarding the internal processing pipeline for Supervised Fine‑Tuning (SFT) when using the ChatCompletion-style JSON format for training data.

{
"messages": [
{
"role": "system",
"content": "You are an ..."
},
{
"role": "user",
"content": "Please tell me about the paper "K205.........""
},
{
"role": "assistant",
"content": "The paper titled "K205R specific nanobody...""
}
]
}

How does the framework internally transform the ChatCompletion training data format into a model ready input. And how is the loss computed