Supervised Fine-Tuning / SFT

Definition

Supervised fine-tuning is a post-training method that trains a model on curated input-output examples or trajectories so its behavior moves toward a target task, style, or workflow.

Current Synthesis

In this episode, SFT is the practical default for individual builders and application companies that want a small model to improve in a specific domain. It can function as a form of Model Distillation / 模型蒸馏 when examples are produced or shaped by a stronger teacher model, but its success depends on target alignment, data quality, and evaluation rather than on the training label alone.

Key Claims

• SFT is often more realistic than RL for low-budget, target-specific model improvement.
• SFT can transfer part of a stronger teacher model’s behavior into a smaller student model when the examples match the desired target.
• SFT does not guarantee general capability improvement; it can improve one benchmark while degrading another.
• The method’s practical difficulty is data construction and review, not only running the training job.
• SFT remains bounded by the base model’s capacity and by whether the selected examples actually shift the learned behavior.

Evidence

Method choice

• 一个人、两周、数百美元，如何训出登顶 Hugging Face 的模型 | 对谈研究员逯雨鑫 says the guest chose SFT because it was more practical for an individual project than RL or more complex post-training methods.

Distillation role

• 一个人、两周、数百美元，如何训出登顶 Hugging Face 的模型 | 对谈研究员逯雨鑫 accepts the host’s summary that stronger teacher-model data can be used through SFT to distill a small model for a narrow domain.

Evaluation loop

• 一个人、两周、数百美元，如何训出登顶 Hugging Face 的模型 | 对谈研究员逯雨鑫 links SFT to benchmark iteration: train, evaluate, inspect failures, change data, and repeat.

Counterevidence & Qualifications

• The source does not provide a full SFT tutorial, hyperparameter recipe, or reproducible training run.
• SFT can overfit a benchmark or pull behavior away from other capabilities if the target and evaluation are too narrow.
• Stronger post-training methods may be necessary when rewards, tool feedback, or long-horizon behavior cannot be captured well by supervised examples.

What Changed

• Created a canonical SFT concept so post-training pages can link to the method without treating it as a one-off abbreviation.

Related Concepts

• Low-Cost Model Post-Training / 低成本模型后训 - application context where SFT is the default low-budget method.
• Model Distillation / 模型蒸馏 - teacher-student transfer pattern that can use SFT data.
• Data-First Post-Training / 数据优先后训 - data discipline that determines SFT usefulness.
• Model Post-Training Bottleneck - broader problem SFT only partially solves.
• Agent Post-Training - harder setting where SFT may need real trajectories and tool feedback.
• AI Verification - evaluation requirement after fine-tuning.