liveBook · Manning

Language model

As explained in Machine Learning System Design

A large language model is a machine learning model trained on large text corpora to generate or transform language and perform related language tasks.

A language model used for dialogue can make monitoring harder than a system expected to return one fixed answer. Repeated prompts to ChatGPT produced different but similar replies; earlier observations also found answers that contradicted one another. The practical question is not whether every reply matches, but which replies are acceptable. section 14.1.2

Treat model changes and output checks as separate operational decisions:

  • When a model is retrained or updated, the same input may produce a different output. Check whether the change was needed and whether its effects are acceptable. section 14.1.2
  • Do not assume that collecting more training examples is enough. New samples help when they add information relevant to the problem; near-duplicates and extreme noise may not. section 6.4
  • Choose serving effort to fit the workload. A time-sensitive request and a scheduled batch job have different latency and cost constraints. section 15.0
Figure 14.2 Annoying ChatGPT to ensure it is a fairly nondeterministic model. Although responses are similar, they are not identical.
figure

Four replies to the same ChatGPT prompt illustrate why identical-output checks are a poor fit for this dialogue example. section 14.1.2

How Machine Learning System Design covers it

How deep it goes
Illustrates response variability and the effects of model updates, but does not explain language-model architecture or training internals.
Examples
Shows four different ChatGPT replies to the same input.
What you can do after
Evaluate/understand it by judging whether varying dialogue responses are acceptable rather than expecting identical replies.
How it connects
Connects variable responses to probabilistic generation, model monitoring, online training, and scheduled retraining.

Also explained in

Compare how 67 books explain language model →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Machine Learning System Design. The section links point to the parts of the book it draws on.