Several thousand to several tens of thousands of high-quality samples are usually enough for supervised fine-tuning (SFT, training a model with “question–standard answer” examples) to lock down output formats and workflows, but they will not close the model’s knowledge gaps. Making a model more obedient does not mean it knows more.
What this is
SFT mainly teaches a model “how to say it,” not “what to know.” It works best for tasks with clear boundaries, stable answer formats, and explicit evaluation criteria—for example, ensuring customer service replies follow company standards or keeping research reports in a consistent structure.
Data preparation usually has three stages: humans write a small set of seed examples, a stronger model expands them in bulk, and rules plus human reviewers filter the results. Real business logs should first be stripped of private information, distilled into fictional tasks, and tested in an isolated environment to assess question difficulty, tool use, and final outcomes.
Rejection sampling fine-tuning (RFT, having a model answer the same question multiple times and retaining only answers that pass verification) can reduce erroneous samples. Distillation (transferring the capabilities of a strong model to a smaller, cheaper model) can lower costs: retaining only conclusions improves efficiency, while retaining the full reasoning process may help a small model recover roughly 70% to 80% of the strong model’s capabilities—but errors are inherited along with them.
Industry view
We believe the article’s most valuable judgment is this: SFT’s ceiling is often determined by data engineering rather than the training algorithm. The clearer the task, the more cost-effective a small, carefully refined dataset becomes; the more open-ended the business problem, the more it relies on external knowledge bases, search, or tool calls to supply factual grounding.
The risks are equally clear. Distillation replicates the teacher model’s systematic errors. A small model trained only on final answers may lose the reliability required when calling tools. SFT may also reinforce stereotypes embedded in the examples. Therefore, “the model has been fine-tuned” cannot be treated as equivalent to “it is ready for unsupervised production deployment.” Continuous evaluation remains a necessary cost.
Impact on regular people
For enterprise IT: When procuring fine-tuning services, organizations should make data quality, error rates, and rollback mechanisms acceptance criteria rather than comparing sample counts and the price of a single training run.
For individual careers: Professional expertise still comes from human judgment and real business knowledge. Fine-tuning is better suited to turning existing capabilities into consistent outputs; it cannot replace learning a new industry, verifying facts, or taking responsibility for decisions.
For the consumer market: Once chat models produce more structured answers, product experiences will become more consistent. However, a professional tone can also conceal knowledge errors, so consumers and enterprises still need to preserve source citations and review checkpoints.