← Back to Documentation

Creating Models

Build and publish your own classification models

Model Development Workflow

1. Train

2. Test

3. Package

4. Publish

Model Configuration

Create a veriafy.yaml file in your model directory:

# veriafy.yaml
name: my-org/fraud-detector
version: 1.0.0
description: Detects fraudulent documents and invoices

# Supported file types
extractors:
  - document  # PDF, DOCX
  - image     # JPG, PNG, WebP

# Classification categories
categories:
  - name: legitimate
    description: Authentic document
  - name: fraudulent
    description: Potentially fake or manipulated
  - name: uncertain
    description: Requires manual review

# Default thresholds
thresholds:
  block: 0.9      # High confidence fraud
  flag: 0.7       # Needs review
  allow: 0.3      # Low confidence fraud

# Model architecture
architecture:
  type: classifier
  base:.veriafy-vector-v2
  hidden_layers: [512, 256, 128]
  activation: relu
  dropout: 0.3

# Training configuration
training:
  epochs: 100
  batch_size: 64
  learning_rate: 0.001
  optimizer: adam
  early_stopping:
    patience: 10
    min_delta: 0.001

# Metadata
license: MIT
author: Your Organization
tags:
  - fraud
  - document
  - compliance

Training Your Model

# Prepare training data
veriafy dataset init ./training-data
veriafy dataset add ./legitimate --label legitimate
veriafy dataset add ./fraudulent --label fraudulent

# Extract vectors from training files
veriafy dataset vectorize ./training-data

# Train the model
veriafy train ./veriafy.yaml --data ./training-data

# Evaluate on test set
veriafy eval ./my-model --data ./test-data

# Output:
# Accuracy: 97.3%
# Precision: 96.8%
# Recall: 97.9%
# F1 Score: 97.3%

Testing Locally

# Test single file
veriafy classify document.pdf --model ./my-model

# Test batch
veriafy classify ./test-files --model ./my-model --output results.json

# Run inference server locally
veriafy serve ./my-model --port 8080

# Test via API
curl -X POST http://localhost:8080/classify \
  -F "file=@document.pdf"

Packaging

# Package model for distribution
veriafy package ./my-model

# Output:
# Created: my-org-fraud-detector-1.0.0.veriafy
# Size: 45.2 MB
# SHA256: abc123...

# Test the package
veriafy run my-org-fraud-detector-1.0.0.veriafy --test

The package includes the model weights, configuration, and metadata in a single portable file.

Publishing to Marketplace

# Login to Veriafy
veriafy login

# Publish model
veriafy push my-org/fraud-detector:1.0.0

# Set visibility
veriafy model set-visibility my-org/fraud-detector public

# Set pricing (optional)
veriafy model set-price my-org/fraud-detector \
  --type subscription \
  --monthly 99

Publishing Checklist

  • Model achieves >90% accuracy on validation set
  • All categories have clear descriptions
  • License is specified and compatible
  • Documentation explains use cases and limitations
  • Test data does not contain PII or sensitive content
  • Model passes automated quality checks

Next Steps

Monetization GuideFine-Tuning Models
Veriafy - Universal File Classification Platform