Term
RLHF
Reinforcement Learning from Human Feedback — training procedure that aligns models with helpful and safe behavior using human preference comparisons.
RLHF — in more detail
RLHF (Reinforcement Learning from Human Feedback) is the standard recipe by which modern chat models like ChatGPT, Claude or Gemini acquire their helpful, friendly and safe response style. The pipeline typically has three stages: first, supervised fine-tuning on high-quality example answers. Second, training a reward model on pairwise human preference comparisons (“answer A is better than answer B”). Third, reinforcement-learning optimization of the language model against this reward model, often using the PPO algorithm.
Example / practical context
Without RLHF, language models behave much more like raw text completers: a question may be answered by another similar question, because training text often contains lists of similar questions. RLHF aligns the model with what humans actually consider a helpful answer. Newer methods such as DPO (Direct Preference Optimization) replace the classic RL step and skip the explicit reward model entirely while keeping the same core idea — learning from preference data.
Delineation from similar terms
RLHF aligns the model with human preferences, while fine-tuning in a narrower sense only learns from new examples. RLAIF (“from AI Feedback”) uses AI-generated comparisons instead of human ones to scale the most expensive step. Constitutional AI is Anthropic’s variant, evaluating responses against an explicit set of principles.
Entdecke mehr
AI Workflows by Keyword: How We Make Recurring Routines Enforceable
A typed keyword triggers a fixed AI routine — and every single step must be committed before the next one appears. Why that's the actual trick.
GlossarDPO (Direct Preference Optimization)
DPO is a fine-tuning method that aligns language models directly to human preference pairs — without a separate reward model and without reinforcement learning.
LexikonFine-Tuning Explained — When, How, With What
When fine-tuning pays off, which methods exist (SFT, DPO, LoRA, QLoRA), and what AMD vs. NVIDIA hardware actually means in practice.