Jargon
- LLM (Large Language Model)
AI model trained on vast text data to generate human-like responses.
- Prompt
Input or question given to the AI to elicit a response.
- Prompt Engineering
Crafting effective prompts to elicit desired AI responses.
- Zero-Shot Prompting
Task given without examples; relies on the model’s training.
- One-Shot Prompting
Task given with one example to guide the response.
- Few-Shot Prompting
Task given with multiple examples to inform the output.
- Fine-Tuning
Adjusting a pre-trained model on specific data for better performance.
- Temperature
Controls randomness in output; lower values = predictable, higher = creative.
- Top-k Sampling
Selects from the top k most likely words to generate responses.
- Context Window
The amount of text the model considers when generating responses.
- Token
Smallest unit of text processed by the model, such as a word or part of one.
- Output Format
Desired structure or style of the AI’s response.
- Response Quality
Measure of accuracy and relevance of the AI’s output.
- Model Bias
Skewed outputs based on training data, reflecting societal biases.
- Inference
Generating predictions from a trained model based on input.
- Batch Size
Number of inputs processed at once during training or inference.
- Model Architecture
Design and structure of the AI model, such as transformers.
- Training Data
Dataset used to train the model, impacting its performance.
- Overfitting
Learning training data too well, leading to poor generalization.
- Transfer Learning
Adapting a pre-trained model for a new task with minimal training.
- Prompt Tokenization
The process of breaking down prompts into individual tokens for processing by the model.
- Pre-training
Initial training phase where the model learns from a large dataset before fine-tuning for specific tasks.
- Loss Function
A mathematical function that measures the difference between the predicted output and the actual output during training.
- Hyperparameters
Configuration settings (like learning rate) that influence model training but are not learned from the data.