AI training data glossary
Last updated: November 9, 2026
These are the terms you will see when comparing training data companies.
Terms
- Annotation (data labeling): Adding labels to raw data, such as bounding boxes, segmentation, transcription or classification, so a model can learn from it.
- Supervised fine-tuning (SFT): Training a model on example prompt-and-response pairs, often written by experts.
- RLHF (reinforcement learning from human feedback): Training a model using people's rankings or ratings of its outputs.
- Preference data: Human comparisons, such as “response A is better than response B,” used in RLHF and in evaluation.
- RL environment: A simulated workspace (code repository, browser, spreadsheet, business app) where a model attempts tasks and receives a reward for success. Vendor examples include Labelbox Horizon, micro1 Realm and Toloka's environments.
- Agent trajectory: A recorded step-by-step sequence of actions an agent or person takes to complete a task, used for training or evaluation.
- Evaluation (evals): Test sets and grading rubrics that measure model quality, often built and scored by domain experts.
- Benchmark / leaderboard: A standardized evaluation with published model scores, such as Mercor's APEX, Snorkel's Terminal-Bench and Surge AI's DAYJOB.
- Red teaming: Deliberately probing a model for unsafe or incorrect behavior.
- Off-the-shelf (OTS) dataset: A dataset that has already been collected and is licensed as-is, as opposed to a custom collection built to your specification.
- Expert network: A vetted pool of specialists (physicians, lawyers, engineers, PhDs) hired for data and evaluation work.
- Inter-annotator agreement: Statistics such as Cohen's kappa or Krippendorff's alpha that measure how consistently different labelers agree.
- Sensor fusion / LiDAR annotation: Labeling 3D point clouds together with camera and radar data, typically for autonomous vehicles and robotics.
- Run rate: Current revenue projected over a year. Gross run rate can include money passed through to contributors.