LLM Training Vocabulary

initial training on vast text data to learn language patterns before task-specific fine-tuning

“Pre-training on internet text gives the model broad world knowledge.”

Origin: Latin prae- `before` + Old French trainer `to draw, drag`

additional training on specific data to adapt a pre-trained model for particular tasks

“Fine-tuning on medical texts improved the model's diagnostic suggestions.”

Origin: Middle English fin `of superior quality` + tune from Greek tonos

reinforcement learning from human feedback—training models using human preference judgments

“RLHF shaped the model's helpfulness by learning from human ratings of responses.”

Origin: Acronym: Reinforcement Learning from Human Feedback

training on labeled examples where correct outputs are provided

“Supervised learning on question-answer pairs teaches the model to respond helpfully.”

Origin: Latin super- `over` + videre `to see` + learning

training where labels are derived from the data itself, like predicting masked words

“Self-supervised learning on next-token prediction requires no human labeling.”

Origin: Greek auto- `self` + Latin super- `over` + videre `to see`

a mathematical measure of how wrong the model's predictions are, minimized during training

“Cross-entropy loss measures how different the predicted token distribution is from the actual next token.”

Origin: Old English los `destruction` + Latin functio `performance`

an optimization algorithm that iteratively adjusts parameters to minimize loss

“Gradient descent slowly nudges billions of parameters toward better predictions.”

Origin: Latin gradus `step` + descendere `to climb down`

the algorithm for computing gradients by propagating errors backward through the network

“Backpropagation calculates how each weight contributed to the prediction error.”

Origin: Latin back + propagare `to extend, spread`

when a model memorizes training data rather than learning generalizable patterns

“Overfitting caused the model to excel on training examples but fail on new ones.”

Origin: Old English ofer `over` + Old Norse fitja `to knit`

techniques to prevent overfitting by constraining model complexity

“Dropout regularization randomly disables neurons during training to improve generalization.”

Origin: Latin regula `rule` + -ization