Training of Large Language Models

A large language model learns by repeatedly predicting text elements called tokens from large collections of examples. During training, it adjusts numerical parameters so that token sequences common in the data receive higher probabilities. It then generates text by using these learned probabilities, rather than by applying a fixed set of hand-written language rules.

From text to training examples

Training text is divided into tokens, which may be words, word fragments or punctuation marks. Tokens are converted into numerical representations that allow the model to process relationships among them.

• In self-supervised learning, the text itself supplies the training signal, commonly through predicting a missing or next token.
• Repeated exposure enables the model to learn statistical regularities involving grammar, meaning, style and associations among concepts.

How the model learns

Most modern LLMs use the Transformer architecture. Its attention mechanism weighs how strongly different tokens in the context relate to one another, while training adjusts parameters to reduce prediction error.

• An optimisation algorithm compares the predicted token probabilities with the actual training tokens and updates the model's parameters.
• Large-scale initial learning is called pre-training; subsequent fine-tuning can adapt the model to instructions, tasks or preferred response behaviour.
• The resulting parameters encode distributed statistical relationships, not a conventional searchable copy of every source document. However, models can sometimes memorise and reproduce training passages.

How learned patterns become text

At generation time, the model receives a prompt and calculates a probability distribution for the next token. A token is selected, added to the context and the process is repeated until the response ends.

• Selection settings can favour the most probable continuation or introduce controlled randomness, causing different outputs from the same prompt.
• Because the objective is plausible continuation, the model may produce fluent but false statements and does not possess human understanding merely by predicting text.

Intellectual property implications

Training corpora may contain copyright-protected expression. Copyright generally protects original expression rather than facts or ideas, while copying, licensing and statutory exceptions are governed by the applicable law.

• Under India's Copyright Act, 1957, the legality of using protected works must be assessed through the copyright owner's rights, licences and the exceptions specified by law.
• Dataset provenance and possible reproduction of memorised passages therefore create distinct questions concerning authorisation, attribution and infringement.

How UPSC asks this