Post-training
Also known as: Post-training / ポストトレーニング / 事後学習
The collective term for all training phases after pre-training — SFT, RLHF, DPO, and other alignment methods — that transform a raw language model into a helpful, safe assistant.
Overview
Post-training converts a raw pre-trained base model into a user-facing assistant. The typical pipeline is: (1) SFT (Supervised Fine-Tuning) to teach instruction-following, then (2) alignment training — RLHF, DPO, or Constitutional AI — to instill helpfulness, harmlessness, and honesty.
Why it matters
A raw pre-trained model excels at next-token prediction but does not follow instructions reliably and may produce harmful content. Post-training is what turns a base model into a useful product like ChatGPT or Claude.
Related Columns
Related Terms
Feel free to contact us
Contact Us