Trace pretraining and model adaptation
Separate broad language-model pretraining from later task and behavior adaptation.
- Compare pretraining, supervised fine-tuning, and preference-based alignment at a high level.
During pretraining, a model learns statistical patterns from a large training corpus, often with an objective such as predicting missing or next tokens. A pretrained model can then be adapted with curated examples, such as supervised instruction-response pairs. Some systems also use preference feedback or other alignment methods to shape responses. These stages affect behavior but do not make a model infallible or remove the need for evaluation.
A small example
stages = ["pretrain on text", "adapt with examples", "evaluate on held-out tasks"]
for stage in stages:
print(stage)pretrain on text adapt with examples evaluate on held-out tasks
Training data quality, provenance, filtering, and representation influence what a model learns. Fine-tuning changes model parameters; prompting changes the input at use time. Retrieval can supply external context without updating model weights. These approaches solve different problems and can be combined.
Key takeaways
Compare pretraining, supervised fine-tuning, and preference-based alignment at a high level.
Measure behavior with realistic examples and inspect important failures.
Lesson quiz
5 questions · pass with 4 correct · up to 50 XP
Passing this quiz completes the lesson and keeps your streak going. Questions you miss come back in review sessions later.
Questions about this lesson
Stuck? Ask. Figured something out? Share it. Explaining is one of the best ways to learn.
Loading posts…