Few-Shot Learning
Học từ vài mẫu (Few-Shot Learning)
Providing a small number of input-output examples in the prompt to teach the model a pattern without changing its weights.
Few-shot learning (in the LLM context) means giving the model a handful of examples — typically 2 to 10 — directly in the prompt before your actual request. The model infers the pattern from the examples and applies it to the new input.
**Example structure**
Input: "The food was terrible and the service was slow." Output: negative
Input: "Absolutely loved the atmosphere and the cocktails." Output: positive
Input: "It was okay, nothing special." Output: [model fills this in]
The model learns "this is a sentiment classification task with three possible outputs" from the examples alone — no fine-tuning needed.
**Why it works**
LLMs are trained on enormous amounts of text that include implicit patterns. Few-shot examples activate in-context learning: the model treats examples as a demonstration of what's expected and extends the pattern. This is sometimes called "in-context learning."
**How many shots?**
Typically 3–8 examples per category or output type. More isn't always better — after about 10 examples you're eating context window without much accuracy gain. For complex tasks, structured chain-of-thought few-shot examples (showing reasoning steps, not just input-output pairs) outperform simple examples.
**Few-shot vs fine-tuning**
Few-shot: no training cost, immediate, takes context window space, lost when context resets. Fine-tuning: one-time training cost, persistent, faster at inference, more reliable for consistent formatting.
**Pitfalls**
Example selection matters enormously — biased examples produce biased outputs. Inconsistent formatting in examples confuses the model. And few-shot examples consume tokens, pushing you closer to the context limit.