How LLMs are Trained
Training LLMs involves:
-
Pre-training – done on huge datasets (often scraped from books, websites, code, and more) to develop general language ability. These datasets are broad but not always carefully curated, so they may contain outdated, biased, or inaccurate information.
-
Fine-tuning – sometimes done to improve performance on specific tasks or to align behaviour with human preferences, often using feedback from real people.
-
Knowledge Cut-off – LLMs are only aware of information available up to their last training date. Some AI tools now integrate 'live search', helping reduce this limitation.
They operate within a context window, meaning they can only remember a limited number of tokens (parts of words) in any conversation.
We'll explore issues like bias, fairness, and control of LLM training data in a later module.