Supervised Fine-Tuning (SFT) serves as the foundational first step in the standard Reinforcement Learning from Human Feedback (RLHF) workflow. While a pre-trained Large Language Model (LLM) possesses broad knowledge, it often lacks the specific instruction-following capabilities, desired output formatting, or conversational style needed for alignment tasks. SFT addresses this gap by adapting the base LLM using a curated dataset of high-quality prompt-response pairs, often called demonstration data.

Think of the SFT phase as providing the model with its initial training on how to behave in the target context. It's not about learning preferences yet; it's about learning the basic rules of the game – how to respond to prompts, what format to use, and perhaps adopting a specific persona (e.g., a helpful assistant).

Initializing the Policy for Reinforcement Learning

The primary technical role of SFT is to produce an initial policy, typically denoted as $\pi_{SFT}$ , which serves as the starting point for the subsequent Reinforcement Learning (RL) phase. The RL algorithm, usually Proximal Policy Optimization (PPO) in this context, doesn't start optimizing from the raw pre-trained model's parameters. Instead, it begins with the parameters of the SFT model.

This initialization offers significant advantages:

Better Starting Point: The $\pi_{SFT}$ is already closer to the desired behavior space than the original pre-trained model. It understands the task format and generates more relevant responses on average.
Improved Sample Efficiency: Because $\pi_{SFT}$ already produces reasonably good outputs, the RL phase requires less exploration to discover high-reward trajectories (prompt-response pairs). The model doesn't waste as much time generating completely irrelevant or poorly formatted text, allowing it to focus on refining outputs based on the learned reward signal more quickly.
Stability: Starting from a model adapted to the target data distribution can lead to more stable RL training compared to starting from a general-purpose model.

The diagram below illustrates how the SFT model fits into the overall process:

Flow diagram illustrating the three stages of RLHF. The SFT model ( $\pi_{SFT}$ ) produced in Stage 1 is used to initialize the RL policy ( $\pi_0$ ) and often serves as the reference policy for the KL penalty in Stage 3. It can also be used to generate initial responses for human comparison data collection in Stage 2.

Establishing a Behavioral Baseline and Reference

Beyond just initializing the weights, the SFT phase establishes a behavioral baseline. The demonstration data teaches the model:

Instruction Following: How to interpret and follow instructions given in the prompt.
Output Formatting: Adhering to specific formats like markdown, code blocks, lists, etc.
Style and Tone: Adopting a desired persona, level of formality, or safety constraints implicitly present in the SFT data.

Crucially, the SFT model ( $\pi_{SFT}$ ) also typically serves as the reference policy for the KL divergence penalty used in the PPO algorithm during the RL phase. The PPO objective function includes a term like $\beta \cdot D_{KL}(\pi_{RL} || \pi_{SFT})$ , where $\pi_{RL}$ is the policy being trained. This KL term penalizes the RL policy for diverging too far from the initial SFT policy on a per-token basis.

Why is this important?

Preserving Capabilities: It helps retain the general language capabilities and knowledge learned during pre-training and SFT, preventing the RL optimization from drastically degrading fluency or coherence while maximizing the reward signal.
Regularization: It acts as a regularizer, preventing the policy from collapsing to simple, repetitive outputs that might exploit the reward model (reward hacking).
Controlled Adaptation: It ensures the RL phase focuses on refining the model based on preferences, rather than fundamentally changing its core behavior in undesirable ways.

Pre-computation for Reward Model Training

While not its primary role, the SFT model is often used in the data generation process for the subsequent reward modeling stage. To collect human preference data, prompts are fed to one or more models (often including the SFT model itself) to generate multiple candidate responses. Human labelers then compare these responses (e.g., choosing the better one in a pair). Therefore, having a competent SFT model facilitates the creation of relevant and diverse candidate responses needed to train an effective reward model.

In summary, the SFT phase is not merely a preliminary step but a foundational component of the RLHF process. It adapts the LLM to the target domain, initializes the policy for efficient RL training, establishes the desired behavioral format and style, and provides the reference point ( $\pi_{SFT}$ ) essential for stable and effective PPO optimization via the KL divergence constraint. A well-executed SFT phase significantly streamlines the subsequent, more complex stages of reward modeling and reinforcement learning.