# How much data do you need to fine-tune an LLM?

Fine-tuning an LLM for a narrow task with LoRA usually takes a few hundred to a few thousand good examples. Broad changes in behavior take tens of thousands or more. Quality and coverage matter more than count.

## Rough ranges

| Goal | Examples |
| --- | --- |
| A narrow task with [LoRA](/glossary/lora), such as a format, a style, extraction or routing | a few hundred to a few thousand |
| Broad behavior across many tasks, or a new domain | tens of thousands or more |

The first row assumes a model that already handles the language and the domain.

## What matters more than count

- **Quality.** The model copies what it sees, mistakes included. A few hundred checked examples often beat thousands of noisy ones.
- **Coverage.** Include the hard cases and edge inputs, in the proportions you expect in use.

[Synthetic data](/glossary/synthetic-data) can fill gaps, if you check it.

## How to tell you have enough

Hold out a test set. Train on half the data, then on all of it, and compare scores on the held-out rows. If the score still climbs, more data will help. If it is flat, work on quality or coverage. With a small set, watch for [overfitting](/glossary/overfitting).

## In Halo

`test_size` carves an eval split out of the data. A row longer than `max_length` is dropped, not truncated, so check the log for how many rows the length filter dropped. See the [supervised fine-tuning guide](/halo/docs/training-methods/sft).

