# gizai:ml/preprocess

A CPU feature pipeline over the native ML data contract. Preprocessing is fit only on training rows and carried alongside the fitted model for inference.

```gizlang
import * as Preprocess from "gizai:ml/preprocess";
```

## Types

### `Standardizer`

```gizlang
type Standardizer = { features: i64; mean: Vec<f64>; scale: Vec<f64>; }
```

Training-set-only mean and population standard deviation per feature.
A constant column has unit scale so it stays zero after centering.

- `features`: `i64`
- `mean`: `Vec<f64>`
- `scale`: `Vec<f64>`

### `Pipeline`

```gizlang
type Pipeline = { transform: Standardizer; model: Model; }
```

A model with its fitted feature transformation. The two belong together at inference.

- `transform`: `M44.Standardizer`
- `model`: `M42.Model`

## Functions

### `standardizer`

```gizlang
function standardizer(data: Data): Standardizer throws string
```

The Standardizer of the training rows of `data`.

Effects: fail.

### `fit`

```gizlang
function fit(task: string, algorithm: string, params: JsonValue, training: Data, seed: u64): Pipeline throws string
```

Fit the scaler on training rows (never on held-out/test rows), then train.

Effects: fail.

### `decode`

```gizlang
function decode(data: bytes): Pipeline throws string
```

Decode and verify the complete scaler/model bundle before inference.

Effects: fail.

## Methods

### `Standardizer.rows`

```gizlang
function Standardizer.rows(self: Standardizer, x: Vec<f64>, n: i64): Vec<f64> throws string
```

Scale one flat row-major matrix, preserving column count and row order.

Effects: fail.

### `Standardizer.meanShiftZ`

```gizlang
function Standardizer.meanShiftZ(self: Standardizer, x: Vec<f64>, rows: i64): Vec<f64> throws string
```

Per-feature absolute shift of a batch's mean, in training standard deviations.
A simple numeric diagnostic, not statistical drift detection or an alerting service.

Effects: fail.

### `Standardizer.data`

```gizlang
function Standardizer.data(self: Standardizer, input: Data): Data throws string
```

Transform features only; the training labels are retained unchanged.

Effects: fail.

### `Pipeline.predict`

```gizlang
function Pipeline.predict(self: Pipeline, row: Vec<f64>): Vec<f64> throws string
```

Predict one unscaled feature row with the original training transformation.

Effects: fail.

### `Pipeline.predictRows`

```gizlang
function Pipeline.predictRows(self: Pipeline, x: Vec<f64>, rows: i64): Vec<f64> throws string
```

Predict a batch of unscaled rows without bypassing the transformation.

Effects: fail.

### `Pipeline.encode`

```gizlang
function Pipeline.encode(self: Pipeline): bytes throws string
```

Portable bundle: its fitted scaler and native model bytes share one artifact.
JSON is used only for the small manifest and base64-encoded model.

Effects: fail.

gizai:ml/preprocess

A CPU feature pipeline over the native ML data contract. Preprocessing is fit only on training rows and carried alongside the fitted model for inference.

gizlang

import * as Preprocess from "gizai:ml/preprocess";

Types

`Standardizer`

gizlang

type Standardizer = { features: i64; mean: Vec<f64>; scale: Vec<f64>; }

Training-set-only mean and population standard deviation per feature. A constant column has unit scale so it stays zero after centering.

features: i64
mean: Vec<f64>
scale: Vec<f64>

`Pipeline`

gizlang

type Pipeline = { transform: Standardizer; model: Model; }

A model with its fitted feature transformation. The two belong together at inference.

transform: M44.Standardizer
model: M42.Model

Functions

`standardizer`

gizlang

function standardizer(data: Data): Standardizer throws string

The Standardizer of the training rows of data.

Effects: fail.

`fit`

gizlang

function fit(task: string, algorithm: string, params: JsonValue, training: Data, seed: u64): Pipeline throws string

Fit the scaler on training rows (never on held-out/test rows), then train.

Effects: fail.

`decode`

gizlang

function decode(data: bytes): Pipeline throws string

Decode and verify the complete scaler/model bundle before inference.

Effects: fail.

Methods

`Standardizer.rows`

gizlang

function Standardizer.rows(self: Standardizer, x: Vec<f64>, n: i64): Vec<f64> throws string

Scale one flat row-major matrix, preserving column count and row order.

Effects: fail.

`Standardizer.meanShiftZ`

gizlang

function Standardizer.meanShiftZ(self: Standardizer, x: Vec<f64>, rows: i64): Vec<f64> throws string

Per-feature absolute shift of a batch's mean, in training standard deviations. A simple numeric diagnostic, not statistical drift detection or an alerting service.

Effects: fail.

`Standardizer.data`

gizlang

function Standardizer.data(self: Standardizer, input: Data): Data throws string

Transform features only; the training labels are retained unchanged.

Effects: fail.

`Pipeline.predict`

gizlang

function Pipeline.predict(self: Pipeline, row: Vec<f64>): Vec<f64> throws string

Predict one unscaled feature row with the original training transformation.

Effects: fail.

`Pipeline.predictRows`

gizlang

function Pipeline.predictRows(self: Pipeline, x: Vec<f64>, rows: i64): Vec<f64> throws string

Predict a batch of unscaled rows without bypassing the transformation.

Effects: fail.

`Pipeline.encode`

gizlang

function Pipeline.encode(self: Pipeline): bytes throws string

Portable bundle: its fitted scaler and native model bytes share one artifact. JSON is used only for the small manifest and base64-encoded model.

Effects: fail.