AI · LESSON 20 / 25 · English
Fine-Tuning & Transfer Learning
Changing model behaviour is different from supplying reference material.
Content reviewed: 22 October 2026 · தமிழில் படிக்க ↗
Understand it in plain language
Fine-tuning updates some or all parameters of a pretrained model using additional training. It can adapt task behaviour or output style, but requires good examples and evaluation. Parameter-efficient methods such as LoRA train a smaller adaptation rather than updating every original weight. RAG supplies context at inference; fine-tuning changes parameters. They can be combined, and neither automatically solves factuality or privacy problems.
How the AI approach works
Transfer learning reuses knowledge from an earlier task or dataset. Fine-tuning is one way to adapt a pretrained model by updating parameters on additional data. Feature extraction with a frozen encoder is another form of reuse. In-context examples guide a request without changing weights, so they are not fine-tuning.
Full fine-tuning updates all selected model parameters and requires memory for training states as well as weights. Parameter-efficient approaches update fewer new or selected parameters. LoRA represents a weight update with lower-rank factors while keeping the base weight frozen in the adaptation setup. The rank, target modules, data, and optimization still affect results. Smaller trainable adapters do not eliminate inference or data costs.
Use fine-tuning when repeated task behavior, domain patterns, or output consistency need adaptation and you have suitable examples. Use retrieval when the main requirement is access to changing or source-citable facts. Begin with an evaluated prompt baseline; RAG and fine-tuning can be combined when their distinct purposes are justified.
Prepare consented, representative examples with consistent formats and labels. Separate sources or conversations across splits. Compare the adapted model with its base model on target tasks, Tamil and English slices, and capabilities you must preserve. Watch for overfitting, memorization, and forgetting. Fine-tuning is not a reliable shortcut for making every answer factual.
Break the idea into steps
- Try a clear prompt and a baseline before deciding to train.
- Prepare representative, permitted examples and a separate evaluation set.
- Compare the adapted model with the base model, including regressions and cost.
Worked example: connect the method to the task
A support team needs the same JSON fields in thousands of responses. Start with schema-constrained output where supported, validation, and a prompt baseline. If consistent errors remain, high-quality input–output examples may justify adaptation. Keep frequently changing refund rules in retrieved policy documents rather than relying on weights to remember the latest version.
Compare base and adapted models on unseen requests, malformed inputs, English, Tamil, and mixed messages. A lower training loss alone does not justify replacing the deployed model.
Apply the idea: check your reasoning
Think before revealing the answer
Would daily fine-tuning be the simplest way to answer from a frequently updated timetable?
Reveal the answer and reasoning
Usually retrieval or an authorized timetable query is a more direct way to obtain current facts. Fine-tuning may adapt response behavior, but should not be confused with a live database update.
A common mistake: Fine-tuning is not a reliable method for deleting secrets or guaranteeing the model will never reveal training content.
Before you move on
- Explain the main idea without repeating the technical label.
- Identify the approach, its assumptions, and the evidence needed to trust the result.
- Complete the challenge, then compare your reasoning with the answer.
Sources & a deeper next step
The refresh uses primary papers and official documentation. These optional links are in English; you can understand this lesson without reading them first. Product versions and availability can change.