# Calibration Data in Neural Network Quantization

## What It Is

Calibration data is a small, representative sample of real inputs (typically hundreds to thousands of examples) used during **post-training quantization (PTQ)** to estimate the statistical properties of network activations.

## Why It's Needed

Quantization maps continuous values to a discrete set of levels. While weights are directly available, the ranges of **activations** depend on the input data. Calibration data lets you observe actual activation distributions to determine:

- **Dynamic range** — min/max values per layer or channel
- **Percentiles/clipping thresholds** — e.g