YAML Configuration

AiDotNet supports full YAML-based configuration for AiModelBuilder. Define your entire model training pipeline in a single YAML file — no code changes needed.

Quick Start


optimizer:
  type: "Adam"

regularization:
  type: "NoRegularization"

preprocessing:
  steps:
    - type: "StandardScaler"
    - type: "SimpleImputer"
      params:
        strategy: "Mean"

var builder = new AiModelBuilder<double, Matrix<double>, Vector<double>>("config.yaml");
var result = await builder.BuildAsync();

Schema Autocomplete

For IntelliSense and validation in VS Code, add the schema directive at the top of your YAML file:

This requires the YAML Language Server extension.

Section Types

AiDotNet YAML configuration supports four section patterns:

1. Enum-Based Selection

Select a predefined algorithm by name. Maps to an enum value.

optimizer:
  type: "Adam"

timeSeriesModel:
  type: "ARIMA"

2. Interface-Based (type + params)

Select a concrete implementation by class name and optionally set properties.

regularization:
  type: "NoRegularization"

tokenizer:
  type: "BPETokenizer"
  params:
    vocabSize: 32000

3. POCO Configuration

Set properties directly on a configuration object. No type key needed.

quantization:
  mode: "Int8"
  strategy: "Dynamic"
  granularity: "PerChannel"
  useSymmetricQuantization: true

4. Pipeline (steps)

Define ordered processing steps, each with a type and optional params.

preprocessing:
  steps:
    - type: "StandardScaler"
    - type: "SimpleImputer"
      params:
        strategy: "Mean"

Deployment and Infrastructure

quantization

Model quantization configuration for lower precision inference.

quantization:
  mode: "Int8"
  strategy: "Dynamic"
  granularity: "PerChannel"
  useSymmetricQuantization: true

Property	Type	Description
`mode`	enum (QuantizationMode)	Quantization bit-width: Int8, Int4, Float16, etc.
`strategy`	enum (QuantizationStrategy)	Dynamic or Static quantization
`granularity`	enum (QuantizationGranularity)	PerTensor or PerChannel
`useSymmetricQuantization`	boolean	Use symmetric vs asymmetric ranges

compression

Model compression configuration for reducing model size.

compression:
  enabled: true
  compressionLevel: 5

caching

Model caching configuration for storing loaded models.

caching:
  enabled: true
  maxCacheSize: 100

abTesting

A/B testing configuration for comparing model versions.

abTesting:
  enabled: true

telemetry

Telemetry configuration for tracking inference metrics.

telemetry:
  enabled: true
  trackLatency: true
  trackErrors: true
  samplingRate: 1.0

mixedPrecision

Mixed precision training configuration.

mixedPrecision:
  enabled: true

inferenceOptimizations

Inference optimization configuration (KV caching, batching, speculative decoding).

inferenceOptimizations:
  enableKVCaching: true
  enableBatching: true
  maxBatchSize: 32

memoryManagement

Training memory management configuration.

memoryManagement:
  enableGradientCheckpointing: true

Interface-Based Sections

These sections use type to select a concrete implementation and params to configure it. Property names in params are case-insensitive.

regularization

regularization:
  type: "NoRegularization"

Available types: NoRegularization

fitDetector

fitDetector:
  type: "DefaultFitDetector"

Available types: DefaultFitDetector

fairnessEvaluator

fairnessEvaluator:
  type: "BasicFairnessEvaluator"

Available types: BasicFairnessEvaluator, ComprehensiveFairnessEvaluator, GroupFairnessEvaluator

promptTemplate

promptTemplate:
  type: "InstructionFollowingTemplate"

Available types: InstructionFollowingTemplate

promptOptimizer

promptOptimizer:
  type: "DiscreteSearchOptimizer"

Available types: DiscreteSearchOptimizer

fewShotExampleSelector

fewShotExampleSelector:
  type: "FixedExampleSelector"

Available types: FixedExampleSelector

promptAnalyzer

promptAnalyzer:
  type: "PatternDetectionAnalyzer"

Available types: PatternDetectionAnalyzer

tokenizer

tokenizer:
  type: "BPETokenizer"
  params:
    vocabSize: 32000

trainingMonitor

trainingMonitor:
  type: "TrainingMonitor"

Pipeline Sections

preprocessing

Define ordered preprocessing steps applied to your data before training.

preprocessing:
  steps:
    - type: "StandardScaler"
    - type: "SimpleImputer"
      params:
        strategy: "Mean"

Complete Example

See examples/configs/ai-model-builder.yaml for a comprehensive example covering all section types.

Generating Documentation Programmatically

You can generate a complete reference document and JSON Schema at runtime:

using AiDotNet.Configuration;

// Generate markdown documentation
var markdown = YamlDocsGenerator.Generate();
File.WriteAllText("yaml-config-reference.md", markdown);

// Generate JSON Schema for editor autocomplete
var jsonSchema = YamlJsonSchema.Generate();
File.WriteAllText("aidotnet-config.schema.json", jsonSchema);