Human-in-the-loop (HITL)

A workflow that keeps a person in a model's decision path — reviewing, correcting or rating outputs and feeding that judgment back into training or evaluation. At the frontier, the human increasingly must be a domain expert rather than a generalist. Full definition →

RLHF (Reinforcement Learning from Human Feedback)

A training method where humans rate or compare model outputs, those preferences train a reward model, and the reward model steers the policy toward preferred behavior. Its quality ceiling is set by the quality of the human feedback. Full definition →

Reward model

A model trained on human preference data to score how good a candidate output is, used to guide reinforcement learning of a policy model. Full definition →

SFT (Supervised Fine-Tuning)

Training a model on prompt-response examples so it imitates demonstrated behavior. The model inherits the quality — and the errors — of whoever wrote the demonstrations. Full definition →

Instruction tuning

Supervised fine-tuning on instruction-and-response pairs so a model learns to follow directions across many tasks. Full definition →

Data annotation

Adding labels, judgments or structure to raw data so a model can learn from it; the broad term covering everything from simple tags to expert reasoning. Full definition →

Data labeling

Attaching correct classes or tags to training data; the core input to supervised learning, often used interchangeably with annotation. Full definition →

Model evaluation

Measuring what a model can and can't do by assessing its outputs against a standard of correctness, quality or safety, using automated metrics, human judgment, or both. Full definition →

Benchmark (held-out set)

A curated test set withheld from training, used to measure model capability on a task; only as discriminating as the expertise that built it. Full definition →

Red-teaming

Deliberately probing a model to elicit unsafe, incorrect or policy-violating outputs so failures are found and fixed before deployment. Full definition →

Preference data

Human judgments comparing or ranking model outputs, used to train reward models for RLHF. Full definition →

Gold standard (reference data)

Expert-verified correct answers used to score contributor work and measure quality objectively. Full definition →

Provenance

A traceable record attaching every judgment to a verified contributor, task spec and timestamp — a record, not a recollection. Full definition →

Targetless calibration

Sensor calibration (camera, lidar, radar) performed without physical calibration targets; a patented Deepen AI method underpinning physical-AI data work. About Deepen AI →