SFT

Model trained to predict human preferences (or utility) for candidate outputs; used in RLHF-style pipelines.

Ensuring model behavior matches human goals, norms, and constraints, including reducing harmful or deceptive outputs.

Automated detection/prevention of disallowed outputs (toxicity, self-harm, illegal instruction, etc.).

Systematic differences in model outcomes across groups; arises from data, labels, and deployment context.

Architecture that retrieves relevant documents (e.g., from a vector DB) and conditions generation on them to reduce h...

Retrieval based on embedding similarity rather than keyword overlap, capturing paraphrases and related concepts.

Breaking documents into pieces for retrieval; chunk size/overlap strongly affect RAG quality.

Constraining outputs to retrieved or provided sources, often with citation, to improve factual reliability.