AI Training Glossary

Every concept you'll encounter in AI-training contracts — defined in plain language, with concrete examples. 48 terms. For compensation, use our pay data and its stated sample.

RLHF

Reinforcement Learning from Human Feedback — the technique that made ChatGPT helpful instead of just plausible.

Reward Model

A small model that learns to predict which AI output a human would prefer — used to train the larger model.

Preference Data

Pairs of AI outputs labeled "A is better than B" — the raw fuel for RLHF.

DPO (Direct Preference Optimization)

A simpler alternative to RLHF that skips the reward model and trains directly from preference pairs.

Instruction Tuning

Training a base model to follow human instructions by showing it examples of good responses.

SFT (Supervised Fine-Tuning)

The technical name for instruction tuning — supervised learning on (input, ideal-output) pairs.

Fine-Tuning

Adapting a pretrained model to a specific task or domain by continuing training on new data.

Pre-training

The expensive first stage where a model learns language by predicting next tokens on internet-scale data.

Evaluation (Eval)

Testing AI outputs against quality criteria — the largest single category of AI training work.

Rubric

The structured scoring guide workers use to evaluate AI outputs consistently.

Red Teaming

Adversarially probing AI for safety failures — getting it to produce harmful, biased, or dangerous outputs.

Gold Label / Gold Standard

A reference answer treated as definitively correct — used to grade workers and AI alike.

Inter-Annotator Agreement (IAA)

How often two workers labeling the same item give the same answer — the metric that catches bad rubrics.

Data Annotation

Labeling raw data (images, text, audio) so AI can learn from it — the foundational AI-training task.

Bounding Box Annotation

Drawing rectangles around objects in images — the bread-and-butter of computer-vision training.

Segmentation

Pixel-level image labeling — every pixel gets a class. Slower than bounding boxes, but more precise.

Audio Transcription

Converting speech to text — verbatim or cleaned, multilingual, with optional speaker tags.

AI Voice Acting

Recording your voice for AI training — text-to-speech, accent variety, or emotion-tagged datasets.

Prompt Engineering

Crafting inputs that get specific, high-quality outputs from AI — both a job and a research field.

Chain of Thought (CoT)

Making AI reason step-by-step instead of jumping to an answer — both a prompting technique and training data type.

Few-Shot Learning

Showing the AI a few examples in the prompt so it picks up a new task without retraining.

Zero-Shot Learning

Asking the AI to do something it wasn't explicitly trained on, with no examples — pure instruction-following.

LLM (Large Language Model)

A neural network trained on massive text corpora to predict and generate language — the technology behind ChatGPT.

Foundation Model

A large model pretrained on broad data, capable of being fine-tuned for many downstream tasks.

Transformer

The neural network architecture that powers all modern AI — invented at Google in 2017, named for its self-attention mechanism.

Tokenization

Breaking text into chunks (tokens) so AI can process it — typically subword units, not whole words.

Context Window

How much text an AI can "see" at once — measured in tokens. Currently 100k-2M for frontier models.

RAG (Retrieval-Augmented Generation)

Letting AI look up information at query-time instead of relying only on what it learned during training.

Embedding

A high-dimensional numerical representation of text, image, or audio — used for semantic search and clustering.

Vector Database

A specialized database for storing and searching embeddings — the backbone of RAG systems.

AI Agent / Agentic AI

AI systems that take multi-step actions to accomplish goals — the new frontier of capability and evaluation.

Tool Use

AI calling external functions (search, calculator, code execution) instead of just generating text.

Function Calling

The structured-output mechanism that lets AI invoke developer-defined functions reliably.

AI Alignment

Ensuring AI systems pursue goals that match human intent and values — the hardest open problem in AI.

Safety Tuning

Training AI to refuse harmful requests, avoid bias, and stay within policy.

Constitutional AI

Anthropic's technique for self-supervised alignment — the AI critiques its own outputs against a written "constitution".

Hallucination

When AI confidently produces false information — the #1 failure mode of modern LLMs.

Bias (in AI)

When AI outputs systematically favor or disfavor groups — can be racial, gender, geographic, ideological, or other.

LoRA (Low-Rank Adaptation)

A cheap fine-tuning technique that updates only a small number of parameters — used everywhere in 2026.

PEFT (Parameter-Efficient Fine-Tuning)

Umbrella term for fine-tuning techniques that update only a fraction of model parameters — LoRA is the most common.

Distillation

Training a smaller model to mimic a larger one — captures most of the capability at a fraction of the cost.

Mixture of Experts (MoE)

A model architecture where different sub-networks specialize in different inputs — gets you GPT-4 quality at smaller compute.

Multimodal AI

AI that handles multiple input types — text, images, audio, video — in a single model.

VLM (Vision Language Model)

Models that take images plus text and produce text — "look at this and tell me X".

Diffusion Model

The neural-network architecture behind image generators like DALL-E and Stable Diffusion.

Synthetic Data

Training data generated by another AI rather than collected from humans — cheap and abundant, but with quality risks.

Temperature (Sampling)

A knob that controls AI randomness — low values produce deterministic output, high values produce creative/varied output.

Top-K / Top-P Sampling

Two ways to limit AI to only "high probability" tokens during generation — combat low-quality outputs.