CORE GUIDE
Supervised fine-tuning
Supervised fine-tuning adapts model behavior using curated input–target examples, but it should solve a demonstrated behavior gap rather than replace application design.
Mental model
SFT changes the model's conditional behavior by repeatedly showing examples of desired responses; it is best viewed as behavior adaptation with a dataset contract, not as a way to upload arbitrary product knowledge.
Why it matters
Fine-tuning is tempting when prompts become complicated, but training creates a slower feedback loop and new data-quality risks. It works best when the desired behavior is stable, repeated across many requests and difficult to express reliably through prompting alone. Dynamic facts, permissions and business rules usually belong in retrieval, tools or application code because they must change without retraining.
01
Tie training examples to a stable behavior objective
Define the behavior gap with an evaluation suite first. Curate representative input–target pairs, including difficult slices and negative cases, and keep training and evaluation data separate. Train a candidate, then compare it with the untuned baseline under the same harness. Check regressions outside the target behavior before deciding the adaptation is worth its serving and maintenance cost.
02
Example: stable structured classification style
A product needs thousands of short classifications in a domain-specific label format. Prompting works but requires many examples and still drifts. A carefully curated SFT dataset teaches the stable label behavior, while current policy text remains in retrieval so changes do not require another training run.
Common failure modes
- Using SFT to memorize frequently changing facts or policy.
- Training before defining an evaluation that proves the behavior gap exists.
- Mixing duplicated or model-generated examples across train and test splits.
Engineering heuristics
- Create the evaluation suite before the training dataset.
- Fine-tune stable repeated behavior; retrieve dynamic knowledge.
- Measure regressions and operational cost, not only improvement on the target slice.
Takeaways
- 01SFT is behavior adaptation, not a general database.
- 02Dataset quality and split integrity determine whether gains are believable.
- 03Training should earn its complexity against simpler application-layer alternatives.
Related concepts from the Knowledge Graph
These relationships come from the canonical graph, not a separate Guide taxonomy.