A tutorial dissecting large-model post-training offers a clear path: a few thousand to several million question-answer samples can transform a base model that only continues text into a conversational model. Notably, the ceiling on model capability increasingly depends on data quality rather than parameter count.
What this is
Supervised Fine-Tuning (SFT) exposes a model to large numbers of “instruction-response” samples, teaching it to answer as requested and stop at the appropriate time. The training objective is the same as in pretraining; the difference lies in the data and loss calculation: errors are calculated only for the response portion, while the prompt is masked.
There are two common data formats: Alpaca for single-turn instructions and ShareGPT for multi-turn conversations. Training and inference must also use a consistent chat template; otherwise, the model will encounter formats it never saw during training.
Low-Rank Adaptation (LoRA) completes fine-tuning by training a small number of additional parameters, usually without updating the entire model. This reduces VRAM and storage costs, making it suitable for teams with limited budgets or a need to adjust models quickly.
Industry view
We believe SFT is shifting model competition from “Who has the larger base model?” toward “Who can encode business experience into data?” Human annotation delivers the highest-quality instruction data, but a million-sample dataset may require a budget in the millions of yuan. Self-Instruct generates data far more cheaply: Alpaca’s 52,000 samples, for example, were once produced with API costs in the hundreds of dollars. Distilling a strong model is equally efficient, but it transfers the teacher model’s biases and expression patterns.
The risks are equally clear: cheaply generated data can amplify mistakes, and once an incorrect answer becomes a training target, the model will reproduce it consistently. LoRA is not a free lunch either; performance remains constrained when domain-specific business data is limited or its distribution changes. Incorrect data formats, inconsistent templates, and insufficient sample coverage can all leave a fine-tuned model performing worse than the original. Continuous evaluation on real business tasks is therefore essential before deployment.
Impact on regular people
For enterprise IT: Model purchasing used to focus primarily on parameter count and leaderboard rankings. Today, private data, annotation workflows, and evaluation systems more directly determine return on investment. LoRA lowers the barrier to experimentation, but data governance remains the primary cost.
For individual professionals: Businesses will find it easier to deploy assistants that comply with internal policies and handle customer service, documentation, and workflow tasks. People who can turn professional expertise and real-world experience into high-quality samples will be more competitive.
For the consumer market: Fine-tuned conversational models will provide fewer irrelevant answers and less unsolicited self-questioning. At the same time, training data of unknown provenance may amplify bias or expose private information, so platforms still need to clarify the boundaries of data use.