SFT Basics

SFT is the other core training concept behind the model-adaptation side of this project.

What SFT Means

SFT stands for Supervised Fine-Tuning.

The idea is straightforward:

1. prepare paired input and target-output examples,
2. train the model on those examples,
3. move the model closer to the target domain and output style.

Why This Project Uses SFT

The base model already has general instruction-following ability, but it is not automatically aligned with postal customer-service language, boundaries, or task formats.

SFT is used here to make the model:

1. sound closer to a postal customer-service assistant,
2. respond better on postal-domain tasks,
3. keep output structure more stable for the project workflow.

Relationship Between SFT and RAG

In this project, SFT and RAG are not substitutes.

They solve different parts of the problem:

1. SFT makes the model behave more like the target assistant.
2. RAG gives the model external knowledge support at answer time.

Current Project Training Path

In the current reconstructed version, the SFT path is mainly organized through Apple MLX and LoRA.

That means:

1. the training objective is SFT,
2. the main parameter-update method is LoRA.