A model contains learned numerical patterns. Training adjusts those patterns using data and an objective. Inference runs the trained model to produce a result.
For language models
- Text is divided into tokens, which may be words, word pieces or punctuation.
- The context window is the amount of tokenised information available for one generation.
- A large language model predicts a likely next token repeatedly.
- Parameters are learned values inside the model, not stored company documents.
Three kinds of adaptation
- Prompt and context: change what the model receives for this request.
- Retrieval: find external information and add relevant parts to the context.
- Fine-tuning: further train a model to influence behaviour or task performance.
These approaches solve different problems. Fine-tuning is not a reliable replacement for retrieving current facts.