Drag each answer to line up with its matching prompt
Pretraining
SFT sharpens helpfulness, tone, and instruction adherence.
SFT
Optimizes next token likelihood over broad unlabeled corpora.
Primary data source
Optimizes supervised response quality on instruction answer pairs.
Main behavior shift
Pretraining uses web/books/code scale; SFT uses curated labeled prompts.
Pretraining teaches broad next-token prediction from unlabeled corpora, while SFT teaches instruction-following behavior from labeled prompt-answer data.
Imagine teaching a chef in two phases. First, they taste thousands of dishes and learn general cooking patterns without anyone grading each sentence they say. That is pretraining. Later, a head chef gives specific customer requests and correct model answers, then scores how well the chef follows instructions. That is SFT, short for supervised fine-tuning. The first phase builds broad food knowledge, and the second phase shapes how the chef responds to requests politely and usefully.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Interviewers ask this boundary because teams routinely pull the wrong lever when model quality drops. People hear that both pretraining and SFT use token-level cross-entropy and conclude they are interchangeable. They are not. The loss surface can look similar while the supervision signal, data distribution, and resulting behavior shift are completely different.
A teacher-level answer draws a clean line: pretraining learns broad continuation structure from unlabeled corpora, while SFT learns preferred assistant behavior from curated prompt-response data. Once you hold that line, debugging choices become clearer. Missing knowledge breadth points upstream to pretraining data and scale choices. Weak instruction following, formatting reliability, or tone control points downstream to SFT data quality and rubric design.
From a mentoring perspective, this topic rewards candidates who connect pretraining versus SFT boundary to operating decisions, not just definitions. The mechanism to state clearly is unlabeled continuation learning versus curated instruction-response supervision. A frequent interview failure is using SFT to patch broad knowledge gaps that require corpus-level learning. When you narrate this topic, include the concrete evidence you would inspect: base capability probes, instruction-following evals, and cross-domain factual recall. Then close with the implementation stance: stage-aware triage that routes capability issues to data/compute and behavior issues to SFT rubrics. That sequence sounds practical because it mirrors how training teams actually debug real regressions rather than debating abstractions.
Same objective family, different teaching signal
The biggest misconception is treating objective form as the whole story. Yes, both stages often optimize token prediction. But in pretraining the target token comes from natural continuation in massive raw corpora. In SFT, the target token comes from selected responses judged desirable for assistant behavior. That means gradient pressure is aimed at different outcomes.
Pretraining gradients reward modeling broad language statistics, world knowledge traces, discourse variability, and long-tail structure. SFT gradients reward behavioral policy: compliance with instructions, response shape, refusal style, and user-facing clarity. So when you compare these phases, do not ask only "what loss?" Ask "who chose targets and from which distribution?" That question reveals what the model is truly being taught.
In practice, this section is where interviewers test decision quality. A strong answer links unlabeled continuation learning versus curated instruction-response supervision to one observable symptom and one corrective action. You can cite base capability probes, instruction-following evals, and cross-domain factual recall as the monitoring surface, then explain how the team decides whether to continue, rollback, or retune. Grounding the explanation in measurable signals prevents the conversation from becoming generic theory and shows that you can operate under uncertainty with finite compute budgets.
A useful teaching pattern is to add a concrete scenario: a chat model that sounds polite after SFT yet still misses long-tail facts. After naming the scenario, state the failure boundary (using SFT to patch broad knowledge gaps that require corpus-level learning) and the operational response (stage-aware triage that routes capability issues to data/compute and behavior issues to SFT rubrics). This structure demonstrates ownership thinking: you are not only describing what the concept means, you are showing how to keep a production run safe when this concept becomes the deciding factor.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
| Axis | Pretraining | SFT |
|---|---|---|
| Data signal | Unlabeled continuation text | Labeled prompt-answer pairs |
| Primary goal | Broad capability and knowledge | Instruction-following behavior |
| Scale | Massive corpus | Curated, much smaller set |
Real products, models, and research that use this idea.
- Open-weight pipelines train a base model on web and code corpora, then run instruction SFT to make chat behavior usable.
- OpenAI and Anthropic product assistants rely on post-pretraining supervised alignment data to shape tone and instruction compliance.
What an interviewer would ask next. Try answering before peeking at the approach.
QWhy can two models with similar pretraining loss still differ after SFT?
Focus on instruction-data quality, rubric consistency, and coverage of user intents rather than base loss alone.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Candidates often say SFT teaches new world knowledge at pretraining scale, instead of saying it mostly reshapes response behavior.
60 second bullets to scan on the way to the call.
Objective difference: pretraining vs SFT
Dataset difference: unlabeled vs labeled
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.