Knowledge / Generative AI

Pretraining and Post-Training

The lifecycle from raw data to useful generative-model behavior.

By Siddhant Krishna · Published 2026-10-06 · Updated 2026-10-06

Training Stages

Pretraining learns broad statistical structure.
Supervised fine-tuning teaches specific response behavior.
Preference training shifts outputs toward preferred behavior.
Domain adaptation specializes behavior for a target environment.
Evaluation and monitoring determine whether the resulting system is fit for deployment.

Data

Training data quality influences model behavior at every stage. Filtering, deduplication, provenance, balancing, and contamination analysis are therefore part of model engineering rather than administrative cleanup.

References

Brown et al. (2020), Advances in Neural Information Processing Systems.
https://arxiv.org/abs/2005.14165
Ouyang et al. (2022).
https://arxiv.org/abs/2203.02155

Large Language ModelsToken prediction, scaling, pretraining, contextual representation, decoding, and emergent task capabilities.Alignment and RLHFInstruction following, preference learning, reward modeling, and behavioral alignment.

Training Stages

Data

Get in Touch