Training, inference, and fine-tuning

Home Learn Training, inference, and fine-tuning

In short

Training changes what an AI model learns from examples. Inference is when the trained model answers a new request. Fine-tuning trains it further for a narrower job.

Why it matters

An assistant can read a document for a response without being retrained on that document. That distinction matters when you’re trying to fix a poor answer. Missing information calls for a different remedy from an unreliable format or a repeated behavior problem. Choosing the technique before diagnosing the failure can add work while leaving the original problem in place.

How it works

Three different activities

A parameter is an adjustable numerical setting in a model. Training changes parameters using examples and an objective: a way to measure how well the model is doing during learning. The model’s architecture is the arrangement of its components. The Transformer research describes one influential architecture and its training for language tasks.

Inference applies the learned model to an input. Supplying a prompt, attaching a document, or retrieving a passage can change the information used for that response. Those actions don’t by themselves change the model’s parameters. An application may separately retain data or use it in later training; that’s a distinct question about the provider’s data practices.

Fine-tuning starts from a model that already learned something and adapts it with selected examples. Some methods update the full model; others update a smaller set of added parameters. The LoRA paper describes one approach that keeps pretrained weights fixed while learning an adaptation. The method doesn’t turn poor examples into good instruction.

Teach behavior or supply knowledge?

Retrieval means finding relevant information and supplying it when the model responds. It can help an application use the current version of a handbook. Fine-tuning can help with a persistent behavior or formatting task. Either choice needs a clear purpose and evaluation.

Start with a baseline: the current result you can compare against. Clarify the task, improve the brief, and inspect the sources. Keep evaluation examples separate from training examples, so the check asks whether the learned behavior works beyond material already used to teach it. Compare changes individually where possible. I’d want to know which change helped before adding another layer.

Example

Illustrative example: this scenario is invented.

Daniel teaches adult-learning classes. He wants an assistant to explain a public course handout in plain language. The handout changes when the class schedule changes. Supplying the approved handout for each response addresses that information need.

He also wants answers to follow a consistent pattern: explanation, practice question, then feedback criteria. He can first put that structure in the prompt. If a larger application keeps failing at the same behavior, a developer could investigate additional training using appropriate examples. Daniel’s changing schedule and his desired teaching format are different problems.

What it doesn’t mean

  • A chat that remembers an earlier message isn’t proof of fine-tuning. Conversation history can supply context.
  • Fine-tuning doesn’t guarantee current facts. The relevant information may still need retrieval.
  • Extra training doesn’t remove review. It can learn defects in the examples along with useful patterns.

Try it

Use about ten minutes with this sample course note: “The evening workshop meets on Tuesday. Bring a notebook. The room number hasn’t been assigned.”

Ask an assistant to explain it using explanation, practice question, and feedback criteria. Then change Tuesday to Thursday and supply the revised note. Check whether it uses the new day and admits the missing room. Discuss which needs came from information and which came from instructions.

Keep going

Sources

Share this page